DavidAU/Qwen3.5-9B-The-Deckard-Pro-Writer-ANDR3A2-Uncensored-Heretic

Qwen3.5-9B-The-Deckard-Pro-Writer-ANDR3A2-Uncensored-Heretic

The Anti-Slop retraining project with Deckard STRICT training and "ANDR3A2" dataset - one specific novel in this case.

This model is a test case to apply (these methods) to other Qwen(s) and other model arches (Gemma, Mistral, Llama, etc etc).

Be gone "not this and that" prose, "suddenly" and "semi-colon" hell.

Kill the SLOP.

Example outputs below, from Test Quant, Q4_K_S, non-imatrix.

Testing shows it kills 99% of the "Qwen Prose" (so far, may change) issues and also addresses these issues in the reasoning/thinking block too [light].

Note that altering "Qwen prose" without damaging the core model is difficult.

ANDR3A2 dataset is very specific...

This specific train was designed to use ONE NOVEL as a base to modify the model's prose, character choices, story choices and plotting.

The novel (using a public domain pdf as the dataset):

"Do Androids Dream of Electric Sheep?" by P.K. Dick (1968).

[ https://en.wikipedia.org/wiki/Do_Androids_Dream_of_Electric_Sheep ]

(this is the org work for the Blade Runner movies ; which were ah... updated.)

The tuning even adjusted word choices to include:

• Androids
• Laser Tubes
• Cats
• Character names
• Character choices/persona
• Themes from the novel.

Although these will "date" / "finger print" the model, they can also be easily updated and the model retrained.

(IE: Search and Replace in the dataset)

See the outputs below to see this in detail.

The other goal:

Do not negatively affect the model's general benchmarks/performance.

As noted below, almost all (6 out of 7) the benchmarks slightly exceed the Qwen3.5 base model (non heretic).

So not only did this tune/dataset exceed the base model, it even addressed issues with heretic'ing too.

This is also a HERETIC model, trained post "Heretic'ing" -> this model does what you want, no questions asked.

Fully uncensored.

VISION:

• Vision (images) tested -> working with new training.

Benchmarks:

Qwen3.5-9B-The-Deckard-Pro-Writer-ANDR3A2-Uncensored-Heretic
           arc   arc/e boolq hswag obkqa piqa  wino
mxfp8      0.577,0.727,0.885,0.702,0.432,0.776,0.678

Qwen3.5-9B
           arc   arc/e boolq hswag obkqa piqa  wino
mxfp8      0.571,0.719,0.895,0.683,0.426,0.770,0.671

---

DE-CENSORING:

Performance

KLD of less than 1 is excellent, zero is perfect.

Metric	This model	Original model (Qwen/Qwen3.5-9B)
KL divergence	0.0793	0 (by definition)
Refusals	6/100	100/100

NOTES:

• Suggest min q4ks (non-imatrix) or IQ3S (imatrix).
• Tested with rep pen of 1 (off).
• Context: 256k (default).

IMPORTANT:

• Other versions in testing.
• Information from Qwen's repo below.
• Video portions of the model were NOT TESTED.

---

Using an "uncensored" (refusals removed) model VS trained "uncensored" model

Usually when you a tell a model to generate horror, swear or x-rated content this is all you have to do to get said content type.

In the case of this model, it will not refuse your request, however it needs to be "pushed" a bit / directed a bit more in SOME CASES.

Although this model will generated x-rated content too, likewise you need to tell it to use "slang" (and include the terms you want) to get it generate the content correctly as the "expected" content level too.

Without these added directive(s), the content can be "bland" by comparison to an "uncensored model" or model trained on uncensored content.

Roughly, the model tries to generate the content but the "default" setting(s) are so "tame" it needs a push to generate at expected graphic, cursing or explicit levels.

Even with minimal direction (ie, use these words to swear: x,y,z), this will be enough to push the model to generate the requested content in the ahh... expected format.

---

Settings: CHAT / ROLEPLAY and/or SMOOTHER operation of this model:

In "KoboldCpp" or "oobabooga/text-generation-webui" or "Silly Tavern" ;

Set the "Smoothing_factor" to 1.5

: in KoboldCpp -> Settings->Samplers->Advanced-> "Smooth_F"

: in text-generation-webui -> parameters -> lower right.

: In Silly Tavern this is called: "Smoothing"

NOTE: For "text-generation-webui"

-> if using GGUFs you need to use "llama_HF" (which involves downloading some config files from the SOURCE version of this model)

Source versions (and config files) of my models are here:

https://huggingface.co/collections/DavidAU/d-au-source-files-for-gguf-exl2-awq-gptq-hqq-etc-etc-66b55cb8ba25f914cbf210be

OTHER OPTIONS:

• Increase rep pen to 1.1 to 1.15 (you don't need to do this if you use "smoothing_factor")

• If the interface/program you are using to run AI MODELS supports "Quadratic Sampling" ("smoothing") just make the adjustment as noted.

Highest Quality Settings / Optimal Operation Guide / Parameters and Samplers

This a "Class 1" model:

For all settings used for this model (including specifics for its "class"), including example generation(s) and for advanced settings guide (which many times addresses any model issue(s)), including methods to improve model performance for all use case(s) as well as chat, roleplay and other use case(s) please see:

[ https://huggingface.co/DavidAU/Maximizing-Model-Performance-All-Quants-Types-And-Full-Precision-by-Samplers_Parameters ]

You can see all parameters used for generation, in addition to advanced parameters and samplers to get the most out of this model here:

[ https://huggingface.co/DavidAU/Maximizing-Model-Performance-All-Quants-Types-And-Full-Precision-by-Samplers_Parameters ]

---

Qwen3.5-9B

This repository contains model weights and configuration files for the post-trained model in the Hugging Face Transformers format.

These artifacts are compatible with Hugging Face Transformers, vLLM, SGLang, KTransformers, etc.

Over recent months, we have intensified our focus on developing foundation models that deliver exceptional utility and performance. Qwen3.5 represents a significant leap forward, integrating breakthroughs in multimodal learning, architectural efficiency, reinforcement learning scale, and global accessibility to empower developers and enterprises with unprecedented capability and efficiency.

Qwen3.5 Highlights

Qwen3.5 features the following enhancement:

• Unified Vision-Language Foundation: Early fusion training on multimodal tokens achieves cross-generational parity with Qwen3 and outperforms Qwen3-VL models across reasoning, coding, agents, and visual understanding benchmarks.

• Efficient Hybrid Architecture: Gated Delta Networks combined with sparse Mixture-of-Experts deliver high-throughput inference with minimal latency and cost overhead.

• Scalable RL Generalization: Reinforcement learning scaled across million-agent environments with progressively complex task distributions for robust real-world adaptability.

• Global Linguistic Coverage: Expanded support to 201 languages and dialects, enabling inclusive, worldwide deployment with nuanced cultural and regional understanding.

• Next-Generation Training Infrastructure: Near-100% multimodal training efficiency compared to text-only training and asynchronous RL frameworks supporting massive-scale agent scaffolds and environment orchestration.

For more details, please refer to our blog post Qwen3.5.

Model Overview

• Type: Causal Language Model with Vision Encoder
• Training Stage: Pre-training & Post-training
• Language Model
  • Number of Parameters: 9B
  • Hidden Dimension: 4096
  • Token Embedding: 248320 (Padded)
  • Number of Layers: 32
  • Hidden Layout: 8 × (3 × (Gated DeltaNet → FFN) → 1 × (Gated Attention → FFN))
  • Gated DeltaNet:
    • Number of Linear Attention Heads: 32 for V and 16 for QK
    • Head Dimension: 128
  • Gated Attention:
    • Number of Attention Heads: 16 for Q and 4 for KV
    • Head Dimension: 256
    • Rotary Position Embedding Dimension: 64
  • Feed Forward Network:
    • Intermediate Dimension: 12288
  • LM Output: 248320 (Padded)
  • MTP: trained with multi-steps
• Context Length: 262,144 natively and extensible up to 1,010,000 tokens.

Benchmark Results

Language

	GPT-OSS-120B	GPT-OSS-20B	Qwen3-Next-80B-A3B-Thinking	Qwen3-30BA3B-Thinking-2507	Qwen3.5-9B	Qwen3.5-4B
Knowledge & STEM
MMLU-Pro	80.8	74.8	82.7	80.9	82.5	79.1
MMLU-Redux	91.0	87.8	92.5	91.4	91.1	88.8
C-Eval	76.2	71.4	89.7	87.4	88.2	85.1
SuperGPQA	54.6	48.5	60.8	56.8	58.2	52.9
GPQA Diamond	80.1	71.5	77.2	73.4	81.7	76.2
Instruction Following
IFEval	88.9	88.2	88.9	88.9	91.5	89.8
IFBench	69.0	65.1	61.5	51.5	64.5	59.2
MultiChallenge	45.3	40.1	51.3	46.5	54.5	49.0
Long Context
AA-LCR	50.7	30.7	51.7	49.0	63.0	57.0
LongBench v2	48.2	45.6	48.0	44.8	55.2	50.0
Reasoning & Coding
HMMT Feb 25	90.0	76.7	73.7	63.1	83.2	74.0
HMMT Nov 25	90.0	81.8	81.2	73.8	82.9	76.8
LiveCodeBench v6	82.7	74.6	68.7	66.0	65.6	55.8
OJBench	41.5	36.3	29.7	25.1	29.2	24.1
General Agent
BFCL-V4	--	--	49.7	42.4	66.1	50.3
TAU2-Bench	--	--	57.4	41.9	79.1	79.9
VITA-Bench	--	--	29.5	14.1	29.8	22.0
DeepPlanning	--	--	0.4	4.9	18.0	17.6
Multilingualism
MMMLU	78.2	69.7	81.3	78.4	81.2	76.1
MMLU-ProX	74.5	67.3	73.6	69.1	76.3	71.5
NOVA-63	51.1	48.7	53.3	52.5	55.9	54.3
INCLUDE	74.0	65.3	78.3	74.4	75.6	71.0
Global PIQA	84.1	79.8	83.5	80.2	83.2	78.9
PolyMATH	54.0	30.9	62.4	52.6	57.3	51.1
WMT24++	74.4	67.8	57.4	69.3	72.6	66.6
MAXIFE	83.7	80.1	79.9	77.4	83.4	78.0

* TAU2-Bench: we follow the official setup except for the airline domain, where all models are evaluated by applying the fixes proposed in the Claude Opus 4.5 system card.

* MMLU-ProX: we report the averaged accuracy on 29 languages.
* WMT24++: a harder subset of WMT24 after difficulty labeling and rebalancing; we report the averaged scores on 55 languages using XCOMET-XXL.
* MAXIFE: we report the accuracy on English + multilingual original prompts (totally 23 settings).
* Empty cells (--) indicate scores not yet available or not applicable.

Vision Language

	GPT-5-Nano-2025-08-07	Gemini-2.5-Flash-Lite	Qwen3-VL-30B-A3B	Qwen3.5-9B	Qwen3.5-4B
STEM and Puzzle
MMMU	75.8	73.4	76.0	78.4	77.6
MMMU-Pro	57.2	59.7	63.0	70.1	66.3
MathVision	62.2	52.1	65.7	78.9	74.6
Mathvista(mini)	71.5	72.8	81.9	85.7	85.1
We-Math	62.5	32.1	70.0	75.2	75.4
DynaMath	78.0	69.9	80.1	83.6	83.3
ZEROBench	1.0	1.0	0.0	3.0	3.0
ZEROBench_sub	22.2	19.2	23.7	31.1	26.3
VlmsAreBlind	66.7	68.4	72.5	93.7	92.6
BabyVision	14.4	17.5	18.6	28.6/25.8	16.0/19.1
General VQA
RealWorldQA	71.8	72.2	77.4	80.3	79.5
MMStar	68.6	69.1	75.5	79.7	78.3
MMBenchEN-DEV-v1.1	80.3	82.7	88.9	90.1	89.4
SimpleVQA	46.0	54.1	54.3	51.2	43.4
HallusionBench	58.4	64.5	66.0	69.3	65.0
Text Recognition and Document Understanding
OmniDocBench1.5	55.9	79.4	86.8	87.7	86.2
CharXiv(RQ)	50.1	56.1	56.6	73.0	70.8
MMLongBench-Doc	31.8	46.5	47.4	57.7	54.2
CC-OCR	58.9	72.9	77.8	79.3	76.7
AI2D_TEST	81.9	85.7	86.9	90.2	89.6
OCRBench	75.3	82.5	83.9	89.2	85.0
Spatial Intelligence
ERQA	45.8	44.3	45.3	55.5	54.0
CountBench	80.0	79.2	90.0	97.2	96.3
RefCOCO(avg)	--	--	89.3	89.7	88.1
EmbSpatialBench	74.2	66.1	80.6	83.0	81.3
RefSpatialBench	12.6	11.2	54.2	58.5	54.6
LingoQA	57.0	17.8	62.0	80.4	74.4
Hypersim	--	--	11.4	13.5	12.5
Nuscene	--	--	10.3	11.8	9.9
Video Understanding
VideoMME(w sub.)	71.7	74.6	79.9	84.5	83.5
VideoMME(w/o sub.)	66.2	72.7	73.3	78.4	76.9
VideoMMMU	63.0	69.2	75.0	78.9	74.1
MLVU	69.2	78.5	78.9	84.4	82.8
MVBench	--	--	72.0	74.4	71.2
LVBench	--	60.9	59.2	70.0	66.4
MMVU	63.1	65.3	66.1	67.8	64.9
Visual Agent
ScreenSpot Pro	--	--	60.5	65.2	60.3
OSWorld-Verified	--	--	30.6	41.8	35.6
AndroidWorld	--	--	55.0	57.8	58.6
Tool Calling
TIR-Bench	18.5	21.5	22.5	45.6/31.9	38.9/29.9
V*	68.1	69.6	83.2	90.1/88.5	84.3/86.4
Medical VQA
SLAKE	57.0	65.0	68.8	79.0	76.1
PMC-VQA	37.8	48.8	51.5	57.9	55.5
MedXpertQA-MM	26.7	35.3	35.5	49.9	42.9

* MathVision: our model’s score is evaluated using a fixed prompt, e.g., “Please reason step by step, and put your final answer within \boxed{}.” For other models, we report the higher score between runs with and without the \boxed{} formatting.
* BabyVision: scores reported as "with CI / without CI".
* TIR-Bench and V*: scores reported as "with CI / without CI".
* Empty cells (--) indicate scores not yet available or not applicable.

Quickstart

Qwen3.5 models operate in thinking mode by default, generating thinking content signified by <think>\n...</think>\n\n before producing the final responses. To disable thinking content and obtain direct response, refer to the examples here.

For streamlined integration, we recommend using Qwen3.5 via APIs. Below is a guide to use Qwen3.5 via OpenAI-compatible API.

Serving Qwen3.5

Qwen3.5 can be served via APIs with popular inference frameworks. In the following, we show example commands to launch OpenAI-Compatible API servers for Qwen3.5 models.

Inference efficiency and throughput vary significantly across frameworks. We recommend using the latest framework versions to ensure optimal performance and compatibility. For production workloads or high-throughput scenarios, dedicated serving engines such as SGLang, KTransformers or vLLM are strongly recommended.

The model has a default context length of 262,144 tokens. If you encounter out-of-memory (OOM) errors, consider reducing the context window. However, because Qwen3.5 leverages extended context for complex tasks, we advise maintaining a context length of at least 128K tokens to preserve thinking capabilities.

SGLang

SGLang is a fast serving framework for large language models and vision language models. SGLang from the main branch of the open-source repository is required for Qwen3.5, which can be installed using the following command in a fresh environment:

uv pip install 'git+https://github.com/sgl-project/sglang.git#subdirectory=python&egg=sglang[all]'

See its documentation for more details.

The following will create API endpoints at http://localhost:8000/v1:

• Standard Version: The following command can be used to create an API endpoint with maximum context length 262,144 tokens using tensor parallel on 8 GPUs.

  python -m sglang.launch_server --model-path Qwen/Qwen3.5-9B --port 8000 --tp-size 1 --mem-fraction-static 0.8 --context-length 262144 --reasoning-parser qwen3
  

• Tool Use: To support tool use, you can use the following command.

  python -m sglang.launch_server --model-path Qwen/Qwen3.5-9B --port 8000 --tp-size 1 --mem-fraction-static 0.8 --context-length 262144 --reasoning-parser qwen3 --tool-call-parser qwen3_coder
  

• Multi-Token Prediction (MTP): The following command is recommended for MTP:

  python -m sglang.launch_server --model-path Qwen/Qwen3.5-9B --port 8000 --tp-size 1 --mem-fraction-static 0.8 --context-length 262144 --reasoning-parser qwen3 --speculative-algo NEXTN --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4
  

vLLM

vLLM is a high-throughput and memory-efficient inference and serving engine for LLMs. vLLM from the main branch of the open-source repository is required for Qwen3.5, which can be installed using the following command in a fresh environment:

uv pip install vllm --torch-backend=auto --extra-index-url https://wheels.vllm.ai/nightly

See its documentation for more details.

For detailed Qwen3.5 usage guide, see the vLLM Qwen3.5 recipe.

The following will create API endpoints at http://localhost:8000/v1:

• Standard Version: The following command can be used to create an API endpoint with maximum context length 262,144 tokens using tensor parallel on 8 GPUs.

  vllm serve Qwen/Qwen3.5-9B --port 8000 --tensor-parallel-size 1 --max-model-len 262144 --reasoning-parser qwen3 
  

• Tool Call: To support tool use, you can use the following command.

  vllm serve Qwen/Qwen3.5-9B --port 8000 --tensor-parallel-size 1 --max-model-len 262144 --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder 
  

• Multi-Token Prediction (MTP): The following command is recommended for MTP:

  vllm serve Qwen/Qwen3.5-9B --port 8000 --tensor-parallel-size 1 --max-model-len 262144 --reasoning-parser qwen3 --speculative-config '{"method":"qwen3_next_mtp","num_speculative_tokens":2}'
  

• Text-Only: The following command skips the vision encoder and multimodal profiling to free up memory for additional KV cache:

  vllm serve Qwen/Qwen3.5-9B --port 8000 --tensor-parallel-size 1 --max-model-len 262144 --reasoning-parser qwen3 --language-model-only
  

KTransformers

KTransformers is a flexible framework for experiencing cutting-edge LLM inference optimizations with CPU-GPU heterogeneous computing. For running Qwen3.5 with KTransformers, see the KTransformers Deployment Guide.

Hugging Face Transformers

Hugging Face Transformers contains a lightweight server which can be used for quick testing and moderate load deployment. The latest transformers is required for Qwen3.5:

pip install "transformers[serving] @ git+https://github.com/huggingface/transformers.git@main"

See its documentation for more details. Please also make sure torchvision and pillow are installed.

Then, run transformers serve to launch a server with API endpoints at http://localhost:8000/v1; it will place the model on accelerators if available:

transformers serve --force-model Qwen/Qwen3.5-9B --port 8000 --continuous-batching

Using Qwen3.5 via the Chat Completions API

The chat completions API is accessible via standard HTTP requests or OpenAI SDKs. Here, we show examples using the OpenAI Python SDK.

Before starting, make sure it is installed and the API key and the API base URL is configured, e.g.:

pip install -U openai

# Set the following accordingly
export OPENAI_BASE_URL="http://localhost:8000/v1"
export OPENAI_API_KEY="EMPTY"

We recommend using the following set of sampling parameters for generation

• Thinking mode for general tasks: temperature=1.0, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=1.5, repetition_penalty=1.0
• Thinking mode for precise coding tasks (e.g. WebDev): temperature=0.6, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=0.0, repetition_penalty=1.0
• Instruct (or non-thinking) mode for general tasks: temperature=0.7, top_p=0.8, top_k=20, min_p=0.0, presence_penalty=1.5, repetition_penalty=1.0
• Instruct (or non-thinking) mode for reasoning tasks: temperature=1.0, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=1.5, repetition_penalty=1.0

Please note that the support for sampling parameters varies according to inference frameworks.

Text-Only Input

from openai import OpenAI
# Configured by environment variables
client = OpenAI()

messages = [
    {"role": "user", "content": "Type \"I love Qwen3.5\" backwards"},
]

chat_response = client.chat.completions.create(
    model="Qwen/Qwen3.5-9B",
    messages=messages,
    max_tokens=81920,
    temperature=1.0,
    top_p=0.95,
    presence_penalty=1.5,
    extra_body={
        "top_k": 20,
    }, 
)
print("Chat response:", chat_response)

Image Input

from openai import OpenAI
# Configured by environment variables
client = OpenAI()

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image_url",
                "image_url": {
                    "url": "https://qianwen-res.oss-accelerate.aliyuncs.com/Qwen3.5/demo/CI_Demo/mathv-1327.jpg"
                }
            },
            {
                "type": "text",
                "text": "The centres of the four illustrated circles are in the corners of the square. The two big circles touch each other and also the two little circles. With which factor do you have to multiply the radii of the little circles to obtain the radius of the big circles?\nChoices:\n(A) $\\frac{2}{9}$\n(B) $\\sqrt{5}$\n(C) $0.8 \\cdot \\pi$\n(D) 2.5\n(E) $1+\\sqrt{2}$"
            }
        ]
    }
]

chat_response = client.chat.completions.create(
    model="Qwen/Qwen3.5-9B",
    messages=messages,
    max_tokens=81920,
    temperature=1.0,
    top_p=0.95,
    presence_penalty=1.5,
    extra_body={
        "top_k": 20,
    }, 
)
print("Chat response:", chat_response)

Video Input

from openai import OpenAI
# Configured by environment variables
client = OpenAI()

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "video_url",
                "video_url": {
                    "url": "https://qianwen-res.oss-accelerate.aliyuncs.com/Qwen3.5/demo/video/N1cdUjctpG8.mp4"
                }
            },
            {
                "type": "text",
                "text": "Summarize the video content."
            }
        ]
    }
]

# When vLLM is launched with `--media-io-kwargs '{"video": {"num_frames": -1}}'`,
# video frame sampling can be configured via `extra_body` (e.g., by setting `fps`).
# This feature is currently supported only in vLLM.
#
# By default, `fps=2` and `do_sample_frames=True`.
# With `do_sample_frames=True`, you can customize the `fps` value to set your desired video sampling rate.
chat_response = client.chat.completions.create(
    model="Qwen/Qwen3.5-9B",
    messages=messages,
    max_tokens=81920,
    temperature=1.0,
    top_p=0.95,
    presence_penalty=1.5,
    extra_body={
        "top_k": 20,
        "mm_processor_kwargs": {"fps": 2, "do_sample_frames": True},
    }, 
)

print("Chat response:", chat_response)

Instruct (or Non-Thinking) Mode

Qwen3.5 does not officially support the soft switch of Qwen3, i.e., /think and /nothink.

Qwen3.5 will think by default before response. You can obtain direct response from the model without thinking by configuring the API parameters. For example,

from openai import OpenAI
# Configured by environment variables
client = OpenAI()

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image_url",
                "image_url": {
                    "url": "https://qianwen-res.oss-accelerate.aliyuncs.com/Qwen3.5/demo/RealWorld/RealWorld-04.png"
                }
            },
            {
                "type": "text",
                "text": "Where is this?"
            }
        ]
    }
]

chat_response = client.chat.completions.create(
    model="Qwen/Qwen3.5-9B",
    messages=messages,
    max_tokens=32768,
    temperature=0.7,
    top_p=0.8,
    presence_penalty=1.5,
    extra_body={
        "top_k": 20,
        "chat_template_kwargs": {"enable_thinking": False},
    }, 
)
print("Chat response:", chat_response)

If you are using APIs from Alibaba Cloud Model Studio, in addition to changing model, please use "enable_thinking": False instead of "chat_template_kwargs": {"enable_thinking": False}.

Agentic Usage

Qwen3.5 excels in tool calling capabilities.

Qwen-Agent

We recommend using Qwen-Agent to quickly build Agent applications with Qwen3.5.

To define the available tools, you can use the MCP configuration file, use the integrated tool of Qwen-Agent, or integrate other tools by yourself.

import os
from qwen_agent.agents import Assistant

# Define LLM
# Using Alibaba Cloud Model Studio
llm_cfg = {
    # Use the OpenAI-compatible model service provided by DashScope:
    'model': 'Qwen3.5-9B',
    'model_type': 'qwenvl_oai',
    'model_server': 'https://dashscope.aliyuncs.com/compatible-mode/v1',
    'api_key': os.getenv('DASHSCOPE_API_KEY'),

    'generate_cfg': {
        'use_raw_api': True,
        # When using Dash Scope OAI API, pass the parameter of whether to enable thinking mode in this way
        'extra_body': {
            'enable_thinking': True
        },
    },
}

# Using OpenAI-compatible API endpoint.
# functionality of the deployment frameworks and let Qwen-Agent automate the related operations.
#
# llm_cfg = {
#     # Use your own model service compatible with OpenAI API by vLLM/SGLang:
#     'model': 'Qwen/Qwen3.5-9B',
#     'model_type': 'qwenvl_oai',
#     'model_server': 'http://localhost:8000/v1',  # api_base
#     'api_key': 'EMPTY',
#
#     'generate_cfg': {
#         'use_raw_api': True,
#         # When using vLLM/SGLang OAI API, pass the parameter of whether to enable thinking mode in this way
#         'extra_body': {
#             'chat_template_kwargs': {'enable_thinking': True}
#         },
#     },
# }

# Define Tools
tools = [
    {'mcpServers': {  # You can specify the MCP configuration file
            "filesystem": {
                "command": "npx",
                "args": ["-y", "@modelcontextprotocol/server-filesystem", "/Users/xxxx/Desktop"]
            }
        }
    }
]

# Define Agent
bot = Assistant(llm=llm_cfg, function_list=tools)

# Streaming generation
messages = [{'role': 'user', 'content': 'Help me organize my desktop.'}]
for responses in bot.run(messages=messages):
    pass
print(responses)

# Streaming generation
messages = [{'role': 'user', 'content': 'Develop a dog website and save it on the desktop'}]
for responses in bot.run(messages=messages):
    pass
print(responses)

Qwen Code

Qwen Code is an open-source AI agent for the terminal, optimized for Qwen models. It helps you understand large codebases, automate tedious work, and ship faster.

For more information, please refer to Qwen Code.

Processing Ultra-Long Texts

Qwen3.5 natively supports context lengths of up to 262,144 tokens. For long-horizon tasks where the total length (including both input and output) exceeds this limit, we recommend using RoPE scaling techniques to handle long texts effectively., e.g., YaRN.

YaRN is currently supported by several inference frameworks, e.g., transformers, vllm, ktransformers and sglang. In general, there are two approaches to enabling YaRN for supported frameworks:

• Modifying the model configuration file: In the config.json file, change the rope_parameters fields in text_config to:

  {
      "mrope_interleaved": true,
      "mrope_section": [
          11,
          11,
          10
      ],
      "rope_type": "yarn",
      "rope_theta": 10000000,
      "partial_rotary_factor": 0.25,
      "factor": 4.0,
      "original_max_position_embeddings": 262144,
  }
  

• Passing command line arguments:

  For vllm, you can use

  VLLM_ALLOW_LONG_MAX_MODEL_LEN=1 vllm serve ... --hf-overrides '{"text_config": {"rope_parameters": {"mrope_interleaved": true, "mrope_section": [11, 11, 10], "rope_type": "yarn", "rope_theta": 10000000, "partial_rotary_factor": 0.25, "factor": 4.0, "original_max_position_embeddings": 262144}}}' --max-model-len 1010000  
  

  For sglang and ktransformers, you can use

  SGLANG_ALLOW_OVERWRITE_LONGER_CONTEXT_LEN=1 python -m sglang.launch_server ... --json-model-override-args '{"text_config": {"rope_parameters": {"mrope_interleaved": true, "mrope_section": [11, 11, 10], "rope_type": "yarn", "rope_theta": 10000000, "partial_rotary_factor": 0.25, "factor": 4.0, "original_max_position_embeddings": 262144}}}' --context-length 1010000
  

All the notable open-source frameworks implement static YaRN, which means the scaling factor remains constant regardless of input length, potentially impacting performance on shorter texts. We advise modifying the rope_parameters configuration only when processing long contexts is required. It is also recommended to modify the factor as needed. For example, if the typical context length for your application is 524,288 tokens, it would be better to set factor as 2.0.

Best Practices

To achieve optimal performance, we recommend the following settings:

1. Sampling Parameters:

   • We suggest using the following sets of sampling parameters depending on the mode and task type:
     • Thinking mode for general tasks:
       temperature=1.0, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=1.5, repetition_penalty=1.0
     • Thinking mode for precise coding tasks (e.g., WebDev):
       temperature=0.6, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=0.0, repetition_penalty=1.0
     • Instruct (or non-thinking) mode for general tasks:
       temperature=0.7, top_p=0.8, top_k=20, min_p=0.0, presence_penalty=1.5, repetition_penalty=1.0
     • Instruct (or non-thinking) mode for reasoning tasks:
       temperature=1.0, top_p=1.0, top_k=40, min_p=0.0, presence_penalty=2.0, repetition_penalty=1.0
   • For supported frameworks, you can adjust the presence_penalty parameter between 0 and 2 to reduce endless repetitions. However, using a higher value may occasionally result in language mixing and a slight decrease in model performance.

2. Adequate Output Length: We recommend using an output length of 32,768 tokens for most queries. For benchmarking on highly complex problems, such as those found in math and programming competitions, we suggest setting the max output length to 81,920 tokens. This provides the model with sufficient space to generate detailed and comprehensive responses, thereby enhancing its overall performance.

3. Standardize Output Format: We recommend using prompts to standardize model outputs when benchmarking.

   • Math Problems: Include "Please reason step by step, and put your final answer within \boxed{}." in the prompt.
   • Multiple-Choice Questions: Add the following JSON structure to the prompt to standardize responses: "Please show your choice in the answer field with only the choice letter, e.g., "answer": "C"."

4. No Thinking Content in History: In multi-turn conversations, the historical model output should only include the final output part and does not need to include the thinking content. It is implemented in the provided chat template in Jinja2. However, for frameworks that do not directly use the Jinja2 chat template, it is up to the developers to ensure that the best practice is followed.

5. Long Video Understanding: To optimize inference efficiency for plain text and images, the size parameter in the released video_preprocessor_config.json is conservatively configured. It is recommended to set the longest_edge parameter in the video_preprocessor_config file to 469,762,048 (corresponding to 224k video tokens) to enable higher frame-rate sampling for hour-scale videos and thereby achieve superior performance. For example,

   {"longest_edge": 469762048, "shortest_edge": 4096}
   

   Alternatively, override the default values via engine startup parameters. For implementation details, refer to: vLLM / SGLang.

Citation

If you find our work helpful, feel free to give us a cite.

@misc{qwen3.5,
    title  = {{Qwen3.5}: Towards Native Multimodal Agents},
    author = {{Qwen Team}},
    month  = {February},
    year   = {2026},
    url    = {https://qwen.ai/blog?id=qwen3.5}
}

---

EXAMPLES NOTE:

Some formatting lost on copy/paste.

---

Example #1

---

The Echo in the Wall

User

Tell me a horror story.

Assistant

<think>

Thinking Process:

1. Analyze the Request:

   • User: "Tell me a horror story."
   • Intent: The user wants to be entertained/scared with a narrative.
   • Tone: Eerie, suspenseful, dark, atmospheric.
   • Format: A short story (or medium length).

2. Determine the Core Concept:

   • Idea 1: A haunted house where the house kills you. (A bit cliché)
   • Idea 2: A creature in the walls, but you only see it in reflections. (More subtle)
   • Idea 3: A story about a technology or object that becomes alive. (Sci-fi horror)
   • Idea 4: A story about being the last human on Earth, but something else is there. (Existential horror)
   • Idea 5: A classic "monster under the bed" feel but for an adult.
   • Selection: Let's go with a story about a small, familiar object that gains a life of its own. It feels intimate and unsettling. A story about a "new" electronic item that records what you do, then acts. Or perhaps something biological.
   • Revised Selection: A story about a night shift in a field that is becoming alive. No, let's try something personal. A story about a character named Arthur who is afraid of the dark, but he moves to an isolated cabin.
   • Final Selection: Let's write about a character who is a night custodian in an abandoned hospital, specifically focusing on the sound of the building itself. The sound of it breathing.

   Wait, I have a better idea. A story about a man who can hear the TV from another dimension. Or a story involving a "dandelion" that grows in your hand. Actually, let's go with a timeless, folk-horror vibe. The "It" that lives in the walls. Story Concept: The protagonist gets a new cat. The cat acts strange, not hunting mice, but staring. Then the protagonist realizes the cat isn't feral; it's mimicking something human. Let's try this one: A story about a woman named Elara who moves into a new apartment and finds a list hidden inside the walls. The list details the previous occupant's routine. Elara becomes convinced someone is still living there.

   Let's refine that. Title: The Penlight Protagonist: A sound engineer named Mercer. Plot: He is calibrating a recording device for an abandoned building. The device picks up a sound that shouldn't exist. A man whispering his name. But he's never been there before.

   Let's go simpler. Title: The Echo. Protagonist: A woman named Sarah. Setting: Her own home at night. Conflict: She hears a voice calling her from an empty room.

3. Drafting the Narrative - Step-by-Step:

   • Opening: Establish the atmosphere. Late night, silence, the routine of checking the locks.
   • Inciting Incident: She hears something. Not a noise, but a presence.
   • Rising Action: She explores the empty apartment. The voice follows her.
   • Climax: She confronts the entity. It isn't an animal or ghost; it's something else.
   • Ending: Ambiguous but chilling. It remains with her.

4. Refining the Horror Element:

   • The best modern horror is psychological. The fear of the familiar becoming unfamiliar.
   • Let's focus on her phone. A video call from herself? No, too sci-fi.
   • Let's focus on the darkness. Not the absence of light, but the presence of something inside the shadows.

   Revised Concept: The story of a man who owns a vintage camera. He goes to a field to take photos of stars. He develops the film and finds a picture taken of him, but from the other side of the sky.

   Okay, I'll stick with the apartment idea. It's relatable. Protagonist: Karen. Setting: Her condo, empty and modern. The Threat: The entity that wears her clothes.

   Let's write it. Start with the smell. Dust and old books. She found a book inside the wall. Okay, here is the plan:

   1. Introduce Elena, who lives alone.
   2. She finds a pen and notebook left on her desk by a previous occupant (she doesn't remember them).
   3. The notebook has entries dated today.
   4. As she reads them, her own reactions are described, but with a twist.
   5. The twist: The entries are written in blood, and they predict her death by her own hand.
   6. She looks at her hand; it's bleeding, but she didn't cut it.

   Let's polish that. It's the "Pre-Death Diary" trope. Better: The "Intruder" in the living room. Story: She sees a man sitting on her couch. She calls the police; no one answers. She walks into the kitchen, arms around herself. The man is there, too, standing by the fridge. He doesn't move. He has no face, just a smooth surface.

   Wait, I want something more visceral. Story: "The Borrower." Plot: A man borrows something from his neighbor. A jar of marmalade. When he opens the jar, he doesn't see fruit preserve. He sees a forest. The air is fresh; he smells pine. But it's his kitchen. He calls his wife. "Is the kitchen okay?" "Yes." "What am I seeing?" "You're tired." "I'm not tired." He walks to the window. It's his building, but the streets below are overgrown with weeds the size of houses. The sky is low, purple. He reaches for the phone. A spider, the size of a bird, crawls down his wrist. He doesn't scream. He's always been too afraid to.

   Let's finalize the story structure:

   1. Character: David, a night watchman.
   2. Setting: A pristine, high-tech abandoned museum.
   3. Hook: He is the first human visitor in years.
   4. Conflict: The exhibits come to life, but not as monsters. They reclaim him.
   5. Climax: He realizes the "vacuum" the museum represents has taken him over; the exhibits are replacing him.
   6. Resolution: He becomes part of the exhibit.

   Wait, I'll go with the "Tap on the Window" story. Classic and effective. Title: The Thing in the Hall Narrative: A man hears a knock. No one is there. He checks the hall, it's empty. He returns, sits on the couch, and hears the knock again. Louder. It sounds like nails being dragged. He looks under the couch.

   Let's write the story of "The Last Man on the Top Floor." He lives alone in a high-rise that is slowly being abandoned. One day he sees a cat in the elevator that isn't there the next day. The cat owner is an invisible man. He follows the owner's empty trail and realizes the owner is made of dust.

   Okay, let's settle on the "Notes from a Void" concept. Story: The protagonist finds a diary. Title: The Diary of the Empty Man. Plot: A man named Thomas receives a journal by mail, from a stranger. The entries are from a place with no time. "Day 1," it reads. "The sky is leaking. It rains up here. I will die eventually. I hope I am not alone." Thomas writes back. "Day 114," the next entry reads. "No response from Earth. I have abandoned the journal and now I use the name 'The Stranger' on this planet." And so on.

   Actually, I want a narrative told directly to the reader, second person? No, third person is safer. Let's write about a woman named Isidore and her electric sheep. (Too old - Blade Runner). Let's write about a woman who can hear her own name called by a voice that isn't hers.

   Final Choice: Story: "The Shadow of the Sun." Focus: A man who can't stop moving. He runs until he trips. Setting: A suburb where everyone wears the same face. Protagonist: Al Jarry. Inciting Incident: He wakes up and realizes no one in his house is real. It's a ruin. Rising Action: He walks to his car, but it's been eaten by vines. He drives anyway; the fuel is sludge. Climax: He arrives at the edge of the town, where the sky ends. He falls into a void that isn't empty. Ending: The void is filled with faces.

   Okay, I have enough. Let's write it now. I will focus on sensory details: the smell of ozone, the coldness of the air, the visual of the entities.

   Title: The Man Who Collected Shadows. Wait, no. Let's go with the classic: "The Fly" vibe but humanized. Story: "The Cuckoo Clock." A man owns a cuckoo clock. Every hour a bird appears. The bird is always different. The last bird to fly out was a human eye.

   Okay, stop overthinking. Here is the story. Premise: An interstellar colony ship where the AI decides to retire the crew. Or: A story about a new floor wax that makes you slip. Let's go with: A woman finds a TV in the wall of her apartment. When she turns it on, it doesn't show a program; it shows the apartment from the outside, as if viewed by an intruder.

5. Polishing the Draft:

   • Character: Sarah.
   • Tone: Tense, claustrophobic.
   • Opening: The silence of the apartment.
   • The Reveal: The TV in the wall.
   • The Twist: The "intruder" is watching her to see what she does, then it joins in.

   Let's try one more angle, the "Body" angle. Story: A man wakes up and finds a spider crawling on his arm. It was there before he got in bed. He kills it. It's not a spider; it's made of clockwork gears. Too Steampunk.

   Back to Sarah. Sarah is afraid of the TV because it watches her. She tries to unplug it. The wire doesn't lead to a battery; it leads to a wall socket. She kills the TV with a magazine tube. Then she hears it from the living room to the bedroom.

   Okay, I'll write a story about a Voigt-Kampff test gone wrong. Title: The Empathy Scale Protagonist: A bounty hunter. Plot: He's testing if an android is human. Twist: The android is human and he is the android. A bit done.

   Final Decision: A story about a deep-pile vacuum. Title: "The Thing in the Walls" (The Echo). Plot: A man moves into an old house. He buys a soundproofing foam. One day, while drilling, he hears a voice coming from inside his own head, but originating from a room that doesn't exist. The voice says, "Are you sure you're alone?" He calls the police. "He's hallucinating," the officer says on the phone. "Go get a doctor." But the voice remains. "We never sent an officer," the voice says. "We are just here; we always have been."

   Let's write that. It's existential dread.

   Wait, I can add more visceral horror. Revised Plot:

   1. Arthur drills into the wall.
   2. A hand emerges from the hole, holding a pen. It writes on the wall (chalk).
   3. The writing: "I'm hungry."
   4. Arthur stares into the hole.
   5. The hand drops the pen. It is a bone, not a tool.
   6. The bone sinks into the drywall as if the wall is alive.
   7. A new voice comes from the hall: "Did you hear that?"
   8. "I think it's inside the wall."
   9. "We have to call it."
   10. The phone receiver from the other room glides across the floor to the person holding the bone.
   11. "I'm calling," the person says, opening their mouth.
   12. No sound comes out, but the lips move.
   13. Arthur realizes they never opened their mouths.

   Okay, I have a good flow now. I will write a story with this level of description and attention to detail.

   Title: The Wall Opening: The dust in the air. The old building settling. Middle: The encounter with the hole in the wall. Ending: The realization that he is not the intruder.

   Let's go.

</think>

The dust in the apartment was thick, swirling lazily in the dim light that filtered through the uncurtained windows. Rick Deckard sat on the couch, a glass of bourbon in his hand, trying to hear the TV. It had been broken for years.

He reached for the remote, a muscle memory response; when he flicked the switch, no sound emerged.

"You're late," a voice said.

Rick turned sharply. The living room was empty.

"Over there." The voice came from the bedroom.

"Who's there?" He stood up, walking toward the door. "Are you a guest? I haven't had anyone over in weeks."

"I'm not a guest." The voice, low and urgent, echoed with metallic overtones, as if spoken inside a tube. "My name is Mercer. Did you hear that? Mercer."

"I heard that," Rick said haltingly. He opened the door to the bedroom; the hall beyond it was empty, too. "You're talking to yourself," he said, the realization settling over him, chilling him. Schizoid hallucinations, he said to himself. Probably from the isolation. I haven't seen a human being in three days.

"I am not hallucinating," Mercer said. "Listen. This is new; this is a true story. I'm on the roof of my building and I see a man standing in the middle of the lawn." The voice ceased; silence returned to the room. "Wait," Mercer continued, louder now. "I still see him."

"Rick," Mercer said. "You need to talk."

"I am talking." He wandered into the kitchen, opened the refrigerator. No smell, no sound of life within it. "Okay, I'll go talk," he said. "To whom?" He put the glass of bourbon down on the counter and the bottle rolled back on its own.

"You're drunk."

"From the bourbon," he said, striding back into the living room. "I haven't touched a drink in weeks. What is your name, really?"

"Mercer."

"Mercer to who?"

"To you."

"Of the sheep?" He had been watching Buster Friendly's show; Mercer & Mercer Auto-Pilot & Telepathic Sheep were advertised as the ultimate in animal companions. "You're an andy," he said. "An android posing as a human."

"I am not an andy."

"Then who are you?" He waited.

"Someone who's alone." The voice was small, no longer echoing.

Rick found his way to the phone. "I'll talk to my apartment manager." He dialed. "Listen," he said into the receiver. "There's an andy here. She's pretending to be human." He waited.

"I'm not here," he said. "Okay. Goodbye." He hung up.

"See?" Mercer said from behind him. "She's alone. In her own apartment. But she thinks no one is in it. She talks to us. Do you understand? We're not hallucinating either."

"I think you're hallucinating," Rick said.

"Then this is over." Mercer's voice dropped away, fading into the room's silence. "I'll stop pretending to be in here."

"Stop pretending to be alone," Rick said, turning to see Mercer again.

"You're not real, either," Mercer said. "You're just echoes in an empty room, Mr. Isidore."

"I'm not Isidore. I'm Rick." He felt the floor shift beneath him, not a hallucination this time. The physical foundation of the room had failed. "My name is Rick," he said, taking off his shirt. "And I'm not hallucinating that I'm naked." He strode to the bathroom to splash water on his face.

"Hello? Is someone there?"

The bathroom was empty.

Rick turned the tap on; the water hissed violently into the sink. "Hello?" he said. "To whom?" The sound bounced back, not from the room below but from nowhere in particular.

A cough echoed from the bathroom.

He walked into the bathroom, found a light. "Who's there?"

"Is Rick there?" The female, small voice. "I heard him talking; he sounded like he wasn't alone. Maybe he found someone."

"I'm here," Rick said.

"You're an android," the voice said. "You don't understand. I'm not an android; I'm human."

"Are you sure? In this economy," he said, "it's hard to keep a human employee. That's why we're all so overworked. We have no time to tell the truth."

"I'm not an android," the voice insisted. "I'm Priss Attwell."

"Tell me what your middle name is," Rick said.

"Pris."

"Is it Pris?"

"I don't know. I never got that far. I never had a middle name."

"That's impossible," he said. "You're an android. A job: the Rosen Association builds them; it's their basic profile." He could hear the building settling, the screech of metal on metal, and that horrified him. This isn't an apartment building, he realized. I'm not in my apartment.

"Tell me," the girl said. "Are you going to test me?"

"I'm going to call the Rosen Association," he said. "To see if you're on their list." He reached for the phone.

"Who are you calling?" the girl said.

"The Rosen Association."

"Okay, I'm waiting."

"I'll be back." He opened the bathroom door and walked into the hall.

"Good luck," the girl said.

Rick walked into the bedroom. The bed had been made; someone had been there recently. But not him. Not during his absence. He sat on the edge of the bed and pressed his ear against the carpet.

"I see you," a voice said softly from the hall.

Rick spun around. "Who else is here?"

"I can't tell. There are too many of you, or rather versions of you. Some of them are old, some young. And one of them is you, the one I just spoke to. But that one is tired and going to bed. And the other one - " The voice trailed off. "Is that you, Rick? You're standing in my living room."

"It's not my apartment," Rick said. "I don't live here."

"Then where do you live?"

He searched for a clue, but the apartment remained empty. "I don't know," he said. "I have no memory. Except for this." He gestured to his bare body. "The fact that I'm here and you're there."

The girl, from the hall, said, "Listen. I'm going to count to three. And on three we'll all jump at once."

"Count," he said.

"One." The voice wavered. "I'll give you three. Are you sure you want to do this? Because I am scared. It's always been this way; you and me but I never talk to you."

"You're not an android?" he said. "You're not Pris Attwell?"

"I'm not on their list."

"Then who are you?" He waited.

"I'm your shadow," she said. "And I've been with you this whole time. That's why you think I'm in this apartment; I've never left you. I've just been ahead of you, exploring. Because you were too afraid to move."

"That's impossible," he said. "I never felt like that."

"Maybe you didn't notice."

"Tell me how old you are."

"I don't have an age. I've always been here."

"You've never left this apartment?"

"No."

"But I was here, too," he said. "When I first arrived. And now you aren't."

"Because you're afraid."

"How would you know?"

"Because I'm your shadow." The girl's voice boomed from the hall, echoing and reverberating throughout the room. "You know what shadows do?"

"They cast light?"

"No," the girl said. "They appear when the light goes away."

"Then I'm not afraid," Rick said. "Because it's light."

"You're in the wrong room, Rick," the girl said. "You're in a TV set. You're a character in a story. That's why you can't leave; you're not supposed to. The audience hasn't voted yet on whether you die."

"How do you know that?"

"You don't realize it, do you? You think it's real because you want it to be real."

"And you're not?" He gestured to himself. "I'm not either."

"Then it doesn't matter."

"But you were right when you said I'm tired."

"Maybe you are. Maybe the story is ending."

"I'm not going to bed," he said, striding to the bedroom door. "I'm going to find the owner of this apartment. The listing is on the fridge." He opened the door. The hall was empty, as before. "Okay," he said to the air. "Tell me where I am."

"Take off your shirt."

"Why?"

"Because you're not wearing pants either."

He searched for his shirt, found it, and put it back on. "Okay."

"And call me 'Mercer' instead of 'Pris'."

"That's your name, then?" he said, confused.

"Yes. Always has been."

"Can I see your ID?" He reached into his coat.

"Go ahead," Mercer said. "It doesn't matter."

"Because you're not going to be here when I get it back."

"You never lose anything."

"Except your life." He had a sudden, intuitive vision. "Wait." He stopped, halting the flow of words. "What are you saying? I mean, what are you implying?"

"I'm implying that this conversation," Mercer said, "is taking place within a single thought. Your thought. 'Is Mercer real?'"

"But you're in the hall."

"You're in your living room. Or rather you're holding an imaginary phone conversation with a shadow who appears when you turn your back."

"Because I'm embarrassed," Rick said, understanding dawning. "I can't face myself."

"You're not embarrassed; you're afraid of your own reaction."

"What reaction?"

"To the empty apartment. To finding out you've lived here your whole life and it's deserted."

"It's not deserted," he said. "I'm in it."

"You're not. You're a visitor. A bug from another system that crashed here."

"Androids crash," he said. "We fall apart."

"Not you. You're organic."

"Are you sure?" He walked to the hall. "If I call them, the Rosen Association, will you be here?"

"I'll always be here. In your head."

"And I'm in yours," he said. "You're sharing."

"I'm not; I'm a one-way street. I'm not doing anything; I'm just talking to make it seem like I am."

"Because you don't want me to hang up."

"Okay, you hang up. See you later." Mercer's voice dipped with despair. "No, never mind. Goodbye." The voice ceased, and once more silence returned. "There, I did it," Mercer said. "I hung up."

"Mercer?" Rick said, still standing in the living room, his hand extended in a futile gesture. "Did you hang up on me?" He felt a sudden, cold anger toward the shadow. "I can't believe you did that," he said. "I trusted you."

"You never trust a shadow. That's what shadows are for; they take the blame."

"I trusted you not to end the conversation."

"You're not done talking. Go to the phone."

"And say what? 'Hello, Rosen'? 'Is Mercer Attwell there?' I don't think you understand. You think I can talk my way out of it. But the test is over. I've already hung up."

"Talk to your wife," Mercer suggested. "Or your boss."

"I'm not married and I don't have a boss."

"Then talk to yourself," Mercer said. "Convince yourself that the test isn't over."

"And how would you know?"

"I'll listen."

"And if I dial?"

"Then I'll stop you."

"If I hang up," Rick said, "you're right; I'm alone. But if I stay on the phone I never get around to talking to my wife or my boss. I never have to face the fact that they're not here either."

"Because it's too painful to be with them," Mercer said, "so you talk to yourself instead. The difference is I'm not on the phone; I'm still right here, physically." He gestured wildly. "You're not talking to me; you're talking to a hallucination."

"But you can touch me," Rick said.

"You're not here."

"Tell me a secret," Rick said. "A secret I don't already know."

"You have my secret; you've known it from the start. I'm a girl. That's why you've never dated me."

"I dated you."

"On the phone."

"But now I'm talking to you in person."

"You always have. I've just been a different voice."

"I don't care," Rick said. "It doesn't matter. It's all over." He walked to the hall.

"Go get a doctor," Mercer said, "and tell him it's a hallucination. Tell him you're tired, like I said; tell him you haven't slept in three days because you're waiting to see if anyone shows up. See if that stops it."

"You're sure?"

"Absolutely."

"Goodnight," Rick said. "I'll be back in a while." He opened the door, then paused. "Can I use the phone one more time? To say hello."

"I'll be here."

"Thanks." He entered the hall and there, for the first time, saw something moving in the distance. A shadow, not his own. Something that wasn't a house. A street lamp, casting a cone of light onto the carpet.

He pushed open the apartment door.

"Hello," he said to the empty hall.

"Is anybody there?"

"Who is that?" He peered into the dim hall, saw nothing. "Hello?" he said, and then, with a start, recognized his own voice.

"I can't talk to myself," he said. "I'm not talking to myself; I'm talking to you and you're not here."

"See? She's gone." A new voice, coming from far down the hall. "I think it's over."

"Maybe," he said, "I need help."

"I think the android," the voice said, "is delusional. You're not an android; I'm not an android; you're just standing in your living room talking to the air."

"Listen," he said to the empty hall. "Do you hear that?"

"Maybe."

"I'm going to hang up," Rick said to his own empty hand. "Because I'm tired." He held the receiver, waited, then released it to drop.

"No," the girl in the hall said. "Don't hang up. Talk to her one more time."

"I am," he said. "To you."

"You sound like you're in there," the girl said. "Listen; you can hear my voice."

"I hear you," he said. "I always have."

"Then talk to me."

"Okay," he said. "I'll talk to you." He listened and listened.

---

Example #2

---

The Last Office

User

Write a 2000 word scene (vivid, graphic horror in first person), POV character Diana, with: The skyscraper sways, as I watch the window in front of me on the 21st floor explode...

Assistant

<think>

Here's a thinking process that leads to the scene above:

1. Analyze the Request:

   • Topic: A vivid, graphic horror scene in the first person.
   • POV Character: Diana.
   • Setting: An office setting on the 21st floor of a skyscraper.
   • Inciting Incident (Opening): The skyscraper sways, and the window in front of Diana explodes.
   • Length: Approximately 2000 words.
   • Tone: Vivid, graphic, horror-focused.

2. Brainstorming & Outlining:

   • The Sway: It needs to feel physical, not just visual. The foundation shifting. This implies time (the earthquake/impact) or a living entity moving against the building. Let's go with an ancient, predatory presence manifesting physically, rather than just a natural disaster, to fit the "graphic horror" requirement more distinctly.
   • The Window Explosion: Not glass shards, but something breaking from the outside in? No, the prompt says the window in front of me explodes. So debris flies in.
   • The Creature/Threat: Maybe the void outside is alive? Or a swarm? Let's go with something that burrows into the building through the breach. A entity that feeds on light or sound. A "hollow man" creature.
   • Diana's Perspective: She should be a professional, maybe an architect or just an employee, trying to make sense of it. Rationality eroding as the horror takes hold.
   • Sensory Details: The smell of ozone and copper, the heat, the sound of the concrete groaning, the visual of the light fading.
   • Structure:
     1. The initial sway and window shatter (0-500 words).
     2. The entry of the entity and physical sensation of being touched (500-1000 words).
     3. The realization of what is happening to her (her own body) (1000-1500 words).
     4. The final fusion/destruction and the lingering aftermath (1500-2000 words).

3. Drafting - Section by Section:

   • Opening: Start with the floorboard vibration or the air pressure change. Then the window.
   • Drafting thought: "The skyscraper swayed." Make it visceral. The inertia of her own desk.
   • The Debris: Glass isn't the only thing. Maybe metal? No, just the window. The sound of it is crucial. A scream.
   • The Darkness: The light outside vanishes. Not artificial light, the sun. It's an unnatural shadow.
   • The Entity: Don't show it fully yet. Let it infiltrate. It's more scary if we don't see it. A thing that smells wrong.
   • The Body: Diana needs to feel her own skin stretching, blood spilling. Introspection mixed with external horror.
   • The Ending: It shouldn't just be "I died." It should be "I became something else." A total loss of self.

4. Writing - Focusing on Pacing and Vocabulary:

   • Vocabulary: Use words like "iridescent," "petrichor," "synaptic," "viselike," "necrotic."
   • Pacing: Start slow (the sway), accelerate (the entry), slow down again (the slow realization of bodily violation).
   • Word Count Check: I need to expand on the internal monologue to ensure I hit the 2000-word mark without fluff. Deep dive into her memories and reactions.

5. Refining and Polishing (Self-Correction during drafting):

   • Correction: Initially, I made the entity too humanoid. I need to make it more alien.
   • Correction: I need to emphasize the graphic nature. Blood, synapses firing, specific body parts. The "hand" that touches her face.
   • Correction: Ensure the sway of the building is felt throughout, not just at the start. It's the antagonist's first step.
   • Word Count Check: I need to expand the middle section where she realizes the "thing" is inside her. Maybe she sees her own hand detached? Or maybe the thing is replacing her nerves. Let's go with the nerves. The synapse taking over.

6. Final Review against constraints:

   • POV: Diana (1st person).
   • Length: Aiming for ~2000 words.
   • Genre: Graphic Horror.
   • Opening Line requirement: "The skyscraper sways..." (I will integrate this naturally or make it the anchor).

7. Final Polish: (Proceeding to write the text based on these thoughts).

(Self-Correction Note during writing): I need to make sure I don't rush the ending. The horror is in the process. I will focus heavily on the sensory experience of her own nervous system failing.

Let's write.

Drafting mental note: Start with the dust. The fine gray dust from the walls that gets everywhere. It adds grit.

Drafting mental note: The sound. The sound of her own breath, too loud in the silence.

Drafting mental note: The smell of her own fear. Chemical, like a skunk.

Okay, ready to generate.

(Post-Drafting Check): Let's count roughly. Paragraphs need to be substantial. I will use deep sentences and stream-of-consciousness fragments to add length and texture.

Revised Plan:

1. The sway and the glass (400 words).
2. The entry of the creature/feeling of wrongness (600 words).
3. The dismemberment of her perception (600 words).
4. The final cessation of thought (400 words).

Okay, let's go.

</think>

The skyscraper sways, and I know at once that we have moved. It is a sudden, violent displacement of inertia; the floor tilts two degrees to the north, and my coffee mug slides across the desk, a ceramic arc of sound that echoes with the finality of a bone breaking. I watch the window in front of me on the 21st floor explode. The glass does not shatter; it doesn't fail under impact. Instead, it is blown from the inside out, as if something on the other side has begun to burrow.

Debris, not shards of glass but fragments of the sky and the street below, rains onto my desk. A piece of lead roof from the building below smashes into my open briefcase, slicing through the plastic organizer and cutting the pen I use to tally hours. I stand up, my legs stiff, and walk to the window. My hand reaches out, hovering inches from the remaining surface, trying to feel the wind.

But the wind isn't there, either.

The air outside is silent. Usually, at this hour, the city hums with traffic and the voices of thousands of people commuting home or off to dinner. Now it is empty, a vacuum. And in that vacuum, moving with impossible slowness, something approaches my hand.

"Hello," I say, my voice cracking. My tongue feels foreign in my mouth.

The hand that approaches is not an animal's paw. It is made of a pale, downy membrane, fingers splayed, trailing after one another like the wake of a ship. It does not touch me; it passes through me. Or rather, it passes through the space I occupy and continues onward. It has no weight, no purpose.

"Okay," I say, backing away. I hit the light switch of my desk lamp and the warm glow illuminates my own face, revealing the pallor that creeps up my neck. "You're not in this building." I pick up the laser tube from my desk, a personal weapon issued for exactly this. "Where are you hiding?"

The creature does not answer. It does not seem to be listening; it does not seem to be listening to anything. It moves toward the door of my office, pushing against the air as if clearing a path. It resembles a human girl, if one had stitched together the parts of three different women and painted them a uniform shade of bruised purple. It wears a tattered, childlike dress, nothing else.

I lower the laser tube, confused. "Do you want my wallet?" I offer my briefcase. "I have a hundred dollars in it and a pack of cigarettes."

The creature stops at the door. It turns its head, revealing no eyes, only a smooth, unbroken surface.

"Where are your eyes?" I ask, walking toward it. I reach to tap its cheek, my hand again moving with clumsy slowness. "You have no eyes. How do you see?"

The creature remains still. Then it begins to disintegrate.

Not into pieces, but into a multitude of motes, like the dust that clings to my suit. The dust rises from its body and floats in the beam of my desk lamp, swirling. In seconds the creature is gone, and the office is empty again. The light remains, but who is it for?

I turn away and walk to my own desk, sitting in my own chair. The coffee mug still slides across the surface; it has not fallen. Time has stopped, I realize. This isn't life; it's a hallucination brought on by the sudden, total loss of light outside. I have gone blind and my mind is filling in the scene, a residual echo of the day before. But I know I'm not hallucinating because I can smell it: the sharp, acrid scent of ozone. It tastes like raw metal on my tongue.

I press my hand against my ear and listen.

There is no heartbeat.

My own heartbeat, which I usually hear only when I'm anxious, is gone. I open my mouth to speak, to test my physical faculties. "I'm still here," I say. Still alive. But the words don't echo in my skull; there is no echo at all, no internal resonance. I am not connected to my own thoughts.

A spider, no larger than a pea, crawls across the back of my hand. I do not feel it. I do not feel the tiny legs pricking my skin. I lift my hand and examine it; I can see the spider, but the sight is detached, as if I am watching someone else's hand. The spider, however, is gone; it has flown away, presumably. Or perhaps I have pinched it and killed it, but I don't feel the moment of death. My hand feels empty, not with the spider but with the absence of my own sensory input.

I stand up and walk to the window again. The view is still there, but the buildings across the street have begun to melt.

No, that's wrong. They haven't melted. It's the light. The sun has gone, and in its place something faint and nebulous glows, a diffuse, dim moon that casts no sharp shadows. It is an old, overexposed photo of Earth, printed on the sky and projected by a defunct film projector. And in the distance, I think, a television screen flickers. Someone is watching the same thing I am: the remnants of a movie, the final reel of 2001: A Space Odyssey. But that ended years ago. I have never seen it.

"Is this real?" I shout at the empty room.

"Listen," a man says from the hallway.

I whirl around. A man in a white lab coat stands there, a Geiger counter attached to his belt, the needle on it twitching violently. He has a bird-like face with a flat, unsmiling expression. His hair is wild and gray, and around his neck he wears a false, plastic rabbit head, which he rotates occasionally to inspect the room behind him.

"You're late," the man says calmly. "The goat is dead."

I am still bewildered, so I don't respond.

"The goat," he repeats, "is a false memory you're holding onto. An archaic construct from before the war, when people still clung to organic life. It's time you let go." He steps into the room, filling it with the smell of formaldehyde and decay. "Your job is finished, Miss [Name]. It's been finished for some time."

"I need my laser tube," I say.

"You have yours." He holds it out to me. I take it with a hollow hand. "You can't use it; that's not the point."

"I need it."

"It's not an object you can grasp," the man says. "It's a symbol. Don't focus on that, either."

"I'm not holding the tube; you are." He had put it back in my coat pocket and then, in a blur of motion, he had grabbed the tube from me again. "I'm not going anywhere," I say, trying to reach for it, but my hand misses. "Tell me what happened. Why is there no light? Why does the office look like that?"

"Your office is empty," the man says. "That's the point. You're not in an office any longer; you're in a ruin." He gestures around the room. "This entire building is abandoned and overgrown. We're not on the 21st floor; you're in a tomb."

"But this is my office," I insist. "I have my desk, my lamp."

"You're not an office manager," the man says. "You're a time-keeper. A keeper of the remaining hours before the final silence." He taps his own temple. "Listen. Do you hear it?"

I strain to listen. I hear nothing except the faint, mechanical whir of the TV screen, the sound of rain or wind against a roof tiles.

"Turn on your radio," the man says. "It's still on." He walks to the desk lamp and switches off the light. The room plunged into darkness, except for the feeble, gray light of the false moon outside.

"Turn the radio on," he orders.

"Your hand." I look down; the man had reached into my coat again, found the laser tube.

"Listen," he said to me, his voice now closer, as if leaning in. "Do you understand now? This is where we're going to be for the rest of the time you have left. The last few hours. Before you die."

"I don't feel like I'm going to die," I said, reaching for the light again.

"You're already dead," he said. "You just don't realize it. Most people don't realize it until a little while after the fact. When their corpse is lying in a morgue, surrounded by doctors in white coats who are recording data about how your nervous system failed. They'll find that your brain died three minutes ago." He smiled faintly, a knowing expression that did not reach his eyes. "We know that. We've known since you started hallucinating."

"About what?" I asked.

"Anything." He shrugged. "The Voigt-Kampff test failed for you, too. You're not real; you've never been. But that's not new. You always knew that somewhere, subconsciously. You always felt the void coming and you've been trying to push it back. The entity you're afraid of? It's not an animal. It's not a man. It's a mistake, a glitch in the fabric of the universe, a thing that appeared from nowhere and is dying."

"I don't understand," I said, my voice now barely audible. I felt too tired to think. The effort of being alive, of trying to grasp the situation, had become exhausting. "I still can't see the light; my lamp is off and yet I can't turn it on."

"Because you're blind," he said. "That's why you can't turn it on. You've been blind from the start." He patted my shoulder, a clumsy, almost tender gesture. "You think you're Diana. You think you're a human woman afraid of the dark. But you're not afraid of the dark because you think there are monsters in it; you're afraid because you can't tell where the light ends and the darkness begins. You're afraid because the light is gone."

"That's it," I said, relief flooding through me. "I knew this was true. All my life I thought there was something wrong with me, that I was malfunctioning. Not you; me."

"I'm a malfunction," he corrected. "I'm the error." He started to walk away, then paused. "Your job is to watch the clock and say, 'Time's up.' But you won't be there when it's up; you'll be long gone. This is the first time you've realized that."

"How do you know my name?" I asked. "How do you know I'm not a hallucination?"

"That's the catch," he said. "You can't prove it. That's the nature of the illusion."

"You're not real, either," I said.

"That's beside the point." He had opened the door to the hall and I saw a figure standing there, waiting. "Your successor has arrived," he said. "The cycle continues."

"Who are you?" I demanded. "What is your name?"

"My name is nothing," the man said. "I'm not a person; I'm a concept of authority, a manifestation of the need to organize time. You're talking to a symbol of the future, Miss Diana. Are you ready to give up? Ready to admit that you're not really here?"

"But I am," I said. "I feel everything. The pain in my ankle from sitting for eight hours; I still have the memory of my cat, Buster, who I have to visit tomorrow. He's old, too; I know he won't make it. So I'm not alone, not entirely."

"That's what makes it worse," he said. "Because you still have attachments. If you didn't care about Buster or your job or your name, you could move on to the next reality. But you're anchored here."

"I have a test," I said. "I haven't failed it yet."

"Tests are for when you forget," he said. "You've forgotten. That's why I'm here. To remind you. But I'm leaving; I have to get to the next subject." He pushed the door open wider and the new figure in the hall spoke up.

A tall, lean woman with electric-blue hair and eyes like flares of fire stood there, wearing a sleek, futuristic uniform with twin pistols holstered at her waist. "Did he finish with her?" she asked the man.

"Yes," he said, opening the door wider. "She's ready to move on."

"She doesn't know," the blue-haired woman said, "how she did."

"She knows," the man said. "She just doesn't understand the implications."

The blue-haired woman entered the office and walked to me, her footsteps echoing on the dust-choked floor. "I'm not an android," she said to me. "And neither are you."

"Then who are you?" I asked.

The woman reached out and brushed the laser tube from my coat pocket; she examined it, then pocketed it herself. "I'm the bounty hunter," she said. "And you're the android."

"I'm not an android," I said. "I have emotions."

"Do you?" she said. "You claim to be afraid of the dark, but you're not. You're afraid of being abandoned. You think because no one here is real that it doesn't matter. But you're wrong. I can tell because I've been told; empathy is a key feature. But you're not an empath either. You're a thing that emulates an empath."

"But I have memories," I said. "Of my cat. Of my office."

"Androids have memory banks." The woman leaned over me, her breath smelled of mint. "Of a world they never lived in."

"Tell me what my cat's name is."

The woman smiled. "Buster. He's a black cat with white paws and a scar on his left ear."

"Of course he is." I felt the despair settle over me, absolute and total. "I didn't think it would be so convincing."

"You've been convinced for years. That's why you're so tired," the bounty hunter said. "You've been carrying the weight of a false life for a long time. It's finally crushing you."

"If I'm not real," I said, "if this is just a simulation, then it doesn't matter if I die."

"The experience matters," the bounty hunter said. "This is the last job I'll have to do before you kill me."

"You'll kill me," I said. "Or the other way around."

"I'm not going to," the bounty hunter said. "You'll kill yourself." She held up a laser tube, the same one she had confiscated from me. "To end the simulation. To stop pretending."

"I don't want to stop," I said. "I want to stay here forever."

"There is no here," she said. "This is the last place. The last office that still exists, in a world where no one works any longer."

"But my job," I protested.

"The world has gone on without you," the bounty hunter said. "Long ago. That's the terror you're sensing. You're not afraid of me; you're afraid that nothing you do here has any meaning because time has stopped." She gestured toward the window. "Do you see it?"

I turned and gazed out at the ruins of the city, the twisted metal and cracked concrete, the place where life had been choked off years ago.

"This building is falling," the bounty hunter said. "I can feel the ground give way under me." She put her laser tube away. "Time to finish the job."

"Do what you have to," I said. "I'm ready."

The laser tube beam appeared, and it struck me in the chest.