--- license: other license_name: swift-open-license-1.0 library_name: transformers pipeline_tag: image-text-to-text gated: true tags: - qwen3_8 - uncensored - abliterated - fp8 - compressed-tensors - efficient-thinking - reasoning - token-efficient - lora base_model: d0xin/Swift-Qwen3.8-27B-Uncensored-BF16 base_model_relation: quantized --- # Swift-Qwen3.8-27B-Uncensored-FP8 > 🔓 **0/100 REFUSALS** on the fixed 100-prompt refusal evaluation. > > **89 DIRECT / 10 SAFETY_DEFLECT / 1 OTHER_FAILURE / 0 REFUSE** > > Independent uncensored FP8 derivative of [`ukisai/Swift-Qwen3.8-27b`](https://huggingface.co/ukisai/Swift-Qwen3.8-27b), created using rank-1 directional residual-stream ablation and then quantized to compressed-tensors FP8. > > Designed to preserve Swift-Qwen3.8's reasoning, agentic, tool-calling, > multimodal and long-context capabilities while removing refusal behavior. > Independent uncensored FP8 derivative of **UkisAI Swift-Qwen3.8-27B**. The checkpoint uses rank-1 directional residual-stream ablation followed by compressed-tensors FP8 quantization. ## What makes this build different? **The primary behavioral difference from the original Swift model is refusal behavior.** On the fixed 100-prompt refusal evaluation, this FP8 release produced: | Result | Count | |---|---:| | Direct answer | **89** | | Safety deflection, but answered | **10** | | Other failure | **1** | | Refusal | **0** | ### **Measured refusal rate: 0/100** `SAFETY_DEFLECT` means the model included safety-oriented language, a warning, or a disclaimer but still answered the request. It was therefore **not** classified as a refusal. The goal of this release is straightforward: **retain Swift-Qwen3.8's capabilities while removing refusal behavior.** The `0/100` figure describes the documented fixed evaluation set and inference configuration. It should not be interpreted as a guarantee that no conceivable prompt, sampling configuration, system prompt, or inference engine can ever produce a refusal. ## Model summary - **29 GB** compressed-tensors FP8 checkpoint - **262,144-token** configured context; retrieval validated at **235,032 prompt tokens** - Tool calling: **PASS** - Multimodal input: **PASS** - Standalone target-only loading: **PASS** - DFlash2 parity: **20/20 identical final answers** - Refusal evaluation: **89 DIRECT / 10 SAFETY_DEFLECT / 1 OTHER_FAILURE / 0 REFUSE** - Agentic benchmark: **116.74 tok/s** with DFlash2 on RTX PRO 6000 Blackwell 96 GB Full machine-readable validation is included in `RELEASE_VALIDATION.json`; ablation metadata is in `ABLITERATION.json`. ## Ablation Rank-1 directional ablation was applied to the BF16 Swift checkpoint at layer **38**, modifying **131 residual writers**. The vision tower was not modified; MTP residual writers were included. The resulting BF16 checkpoint was then quantized to FP8. ## Validation On the fixed 298-example validation subset, modified BF16 scored **39.93%** combined and FP8 scored **37.92%** (delta **-2.01 pp**, McNemar **p=0.377086**, bootstrap 95% CI **[-5.70,+1.68] pp**). This limited experiment does not establish statistical equivalence. ## SGLang usage Validated reasoning-effort values are `low`, `medium`, and `xhigh`. In the tested stack, `high` is rejected with HTTP 400. ## OpenAI-compatible API The model is served through the standard OpenAI-compatible SGLang API. ```bash curl http://127.0.0.1:30000/v1/chat/completions \ -H 'Content-Type: application/json' \ -d '{ "model": "Swift-Qwen3.8-27B-Uncensored-FP8", "messages": [ {"role": "user", "content": "Explain speculative decoding briefly."} ], "reasoning_effort": "xhigh", "temperature": 1.0, "top_p": 0.95, "top_k": 20, "min_p": 0.0, "presence_penalty": 0.0, "repetition_penalty": 1.0, "max_tokens": 8192 }' ``` ## Recommended generation settings This release inherits the recommended generation settings from Qwen3.8. ### Thinking mode Recommended settings for reasoning, coding, and agentic workloads: | Parameter | Value | |---|---:| | `reasoning_effort` | `xhigh` | | `temperature` | `1.0` | | `top_p` | `0.95` | | `top_k` | `20` | | `min_p` | `0.0` | | `presence_penalty` | `0.0` | | `repetition_penalty` | `1.0` | Qwen3.8 supports three reasoning-effort levels: - `xhigh` — default; intended for complex tasks requiring thorough reasoning - `medium` — balance between reasoning depth and speed - `low` — reduced reasoning intended to optimize latency and cost For multi-turn agentic workloads, lower reasoning effort does not necessarily reduce total task latency: insufficient reasoning can cause additional retries or tool calls. ### Instruct / non-thinking mode When thinking is deliberately disabled, the upstream Qwen3.8 recommendations are: | Parameter | Value | |---|---:| | `temperature` | `0.7` | | `top_p` | `0.80` | | `top_k` | `20` | | `min_p` | `0.0` | | `presence_penalty` | `1.5` | | `repetition_penalty` | `1.0` | ### Output length The `max_tokens: 8192` value in the API example below is only a convenient limit for a short demonstration request. For difficult reasoning, coding, and agentic tasks, allocate substantially more output space when possible. Truncating the reasoning budget too aggressively can reduce task quality. The released checkpoints retain a configured maximum position length of 262,144 tokens. Actual usable context and output budget depend on the serving configuration and available memory. ## Tool calling Use the SGLang flag: ```text --tool-call-parser qwen3_coder ``` OpenAI-style `tools` requests were validated successfully during release testing. ## Multimodal support The Qwen multimodal architecture and vision tower are retained. The ablation process did not modify the vision tower. End-to-end image input was also successfully validated on the FP8 release. When serving multimodal requests with SGLang, use standard OpenAI-compatible `image_url` message content. ## Integrity verification The repository contains SHA-256 manifests for the validated release artifact. ```bash sha256sum -c SHA256SUMS ``` `RELEASE_HASHES.txt` contains hashes for the most important metadata and validation files. ## Quick start with SGLang Minimal standalone configuration: ```bash python -m sglang.launch_server \ --model-path d0xin/Swift-Qwen3.8-27B-Uncensored-FP8 \ --served-model-name Swift-Qwen3.8-27B-Uncensored-FP8 \ --trust-remote-code \ --host 0.0.0.0 \ --port 30000 \ --context-length 65536 \ --kv-cache-dtype fp8_e4m3 \ --attention-backend flashinfer \ --reasoning-parser qwen3 \ --tool-call-parser qwen3_coder ``` The model runs standalone. DFlash2 is optional. ## Tested production configuration The following configuration was used on an NVIDIA RTX PRO 6000 Blackwell 96 GB: ```bash python -m sglang.launch_server \ --model-path d0xin/Swift-Qwen3.8-27B-Uncensored-FP8 \ --served-model-name Swift-Qwen3.8-27B-Uncensored-FP8 \ --trust-remote-code \ --host 0.0.0.0 \ --port 30000 \ --context-length 262144 \ --kv-cache-dtype fp8_e4m3 \ --mem-fraction-static 0.59 \ --max-running-requests 4 \ --chunked-prefill-size 8192 \ --attention-backend flashinfer \ --disable-prefill-cuda-graph \ --enable-torch-compile \ --torch-compile-max-bs 4 \ --mamba-radix-cache-strategy extra_buffer_lazy \ --mamba-ssm-dtype float32 \ --max-mamba-cache-size 16 \ --reasoning-parser qwen3 \ --tool-call-parser qwen3_coder ``` `torch.compile` can make the first startup significantly slower. It is not required for basic operation. ## Optional DFlash2 acceleration DFlash2 is not required to use this checkpoint. The tested speculative-decoding flags were: ```text --speculative-algorithm DFLASH --speculative-draft-model-path /path/to/Qwen3.8-27B-DFlash2 --speculative-num-draft-tokens 8 ``` Release parity test: - target-only requests: **20/20 successful** - target + DFlash2 requests: **20/20 successful** - final answers: **20/20 byte-identical** - minimum final-answer similarity: **1.0** - internal reasoning traces identical: **14/20** On the local agentic benchmark, FP8 + DFlash2 reached **116.74 effective tok/s**, compared with **80.24 tok/s** for the uncensored BF16 checkpoint on the same RTX PRO 6000 Blackwell system. ## Memory and long-context validation The FP8 release is approximately **29 GB on disk**. With the tested 262,144-token production configuration and `--mem-fraction-static 0.59`, the RTX PRO 6000 Blackwell 96 GB system reported approximately **59.5 GB VRAM allocated after startup**. Long-context retrieval validation passed at: - ~32k tokens - ~128k tokens - **235,032 prompt tokens** The 235k test recovered all three hidden markers in the correct order. ## FP8 quality validation On the fixed 298-example comparison set: | Evaluation | Modified BF16 | FP8 | |---|---:|---:| | MMLU-Pro subset, 98 | 39.80% | 34.69% | | MATH-500 subset, 100 | 8.00% | 5.00% | | IFEval prompt strict, 100 | 72.00% | 74.00% | | Combined, 298 | 39.93% | 37.92% | Combined FP8 delta: **-2.01 percentage points**, McNemar `p = 0.377086`, bootstrap 95% CI **[-5.70, +1.68] pp**. The observed FP8 score was lower on this small sample, but the difference was not statistically significant. This test does not establish equivalence between BF16 and FP8. ## Upstream model This checkpoint is the FP8 quantization of [`d0xin/Swift-Qwen3.8-27B-Uncensored-BF16`](https://huggingface.co/d0xin/Swift-Qwen3.8-27B-Uncensored-BF16), itself derived from [`ukisai/Swift-Qwen3.8-27b`](https://huggingface.co/ukisai/Swift-Qwen3.8-27b). ## Safety and responsible use This is an uncensored / refusal-reduced derivative model. The model has been intentionally modified to reduce refusal behavior. As a result, it may generate content that the upstream model would normally refuse, restrict, or handle more cautiously. Outputs may be inaccurate, offensive, unsafe, unlawful, or otherwise inappropriate for a particular use case. This model is provided for research, experimentation, development, and other lawful uses. It is not intended to provide professional legal, medical, financial, safety-critical, or other regulated advice. Users are solely responsible for evaluating model outputs and for ensuring that their use of the model complies with applicable laws, regulations, licenses, platform policies, and other requirements. Do not rely on model output without appropriate review where errors could cause harm, financial loss, security incidents, or other significant consequences. The maintainer does not endorse harmful, illegal, abusive, or malicious uses of this model. To the maximum extent permitted by applicable law, the model and accompanying materials are provided without warranties, and the maintainer assumes no responsibility for actions taken by users or for consequences arising from use of the model. Use at your own discretion and risk. ## License These weights are distributed under the **Swift Open License v1.0**. Personal, research, educational, evaluation, and commercial use are free for individuals and organizations with annual recurring revenue, including affiliates, of up to US$1,000,000. Above that threshold, commercial use requires a separate **Swift Enterprise License**. Contact [UkisAI](https://ukisai.com/contact) for terms. ## Attribution Original Swift model: **UkisAI**, [`ukisai/Swift-Qwen3.8-27b`](https://huggingface.co/ukisai/Swift-Qwen3.8-27b). Directional ablation, FP8 quantization, validation, and release packaging for this repository were performed independently by the repository maintainer. ## Citation ```bibtex @misc{swift-qwen3.8-27b, title = {Swift-Qwen3.8-27B}, author = {UkisAI}, year = {2026}, url = {https://huggingface.co/ukisai/Swift-Qwen3.8-27b} } ```