Instructions to use horiuchinobuyuki/Qwick-3.5-9B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use horiuchinobuyuki/Qwick-3.5-9B with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="horiuchinobuyuki/Qwick-3.5-9B")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("horiuchinobuyuki/Qwick-3.5-9B")
model = AutoModelForMultimodalLM.from_pretrained("horiuchinobuyuki/Qwick-3.5-9B", device_map="auto")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use horiuchinobuyuki/Qwick-3.5-9B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "horiuchinobuyuki/Qwick-3.5-9B"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "horiuchinobuyuki/Qwick-3.5-9B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/horiuchinobuyuki/Qwick-3.5-9B

SGLang

How to use horiuchinobuyuki/Qwick-3.5-9B with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "horiuchinobuyuki/Qwick-3.5-9B" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "horiuchinobuyuki/Qwick-3.5-9B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "horiuchinobuyuki/Qwick-3.5-9B" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "horiuchinobuyuki/Qwick-3.5-9B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use horiuchinobuyuki/Qwick-3.5-9B with Docker Model Runner:
```
docker model run hf.co/horiuchinobuyuki/Qwick-3.5-9B
```

Qwick-3.5-9B / EVALUATION.md

horiuchinobuyuki

Upload folder using huggingface_hub

9281e75 verified 11 days ago

preview code

Raw

History Blame Contribute Delete

8.03 kB

Evaluation

The tables below come from local matched runs of Qwen/Qwen3.5-9B@c202236235762e1c871ad0ccb60c8ee5ba337b9a and Qwick-3.5-9B. Each comparison used the same prompts, grader, and decoding profile for both models. Upstream model-card scores use different harnesses and belong to a separate comparison.

raw correct counts every correct answer. budget correct also requires finish_reason != length. Completion length covers thinking and the final answer.

Public benchmark comparisons

Benchmark / primary metric	Qwen3.5-9B	Qwick-3.5-9B	Delta	Token ratio	Profile
MMLU-Pro last-match, 12,032	9,792 (81.383%)	9,760 (81.117%)	-0.266 pp	0.604890	A
GPQA-Diamond, 198	154 (77.778%)	158 (79.798%)	+2.020 pp	0.710227	A
HMMT public 60 × 4 seeds, budget	187/240	172/240	-6.250 pp	0.679427	D
IFEval prompt strict, 541	89.279%	89.649%	+0.370 pp	0.772180	B
IFEval instruction strict	92.206%	92.086%	-0.120 pp	0.772180	B
MMMU validation official-compatible budget, 900	664 (73.778%)	671 (74.556%)	+0.778 pp	0.740937	V
MMMU validation strict budget, 900	662 (73.556%)	667 (74.111%)	n/a	0.740937	V
JMMLU common-clean raw, 987	868 (87.943%)	858 (86.930%)	-1.013 pp	0.4595	A
LiveCodeBench v6 raw, 1,055	766 (72.607%)	785 (74.408%)	+1.801 pp	0.487927	C
LiveCodeBench v6 budget, 1,055	702 (66.540%)	779 (73.839%)	+7.299 pp	0.487927	C

Decoding and grading profiles

A (general thinking): temperature 1.0, top-p 0.95, top-k 20, min-p 0, presence penalty 1.5, repetition penalty 1.0, maximum 32,768 new tokens. MMLU-Pro used five-shot mmlu_pro_fulltrace, with last-match extraction as primary and first-match as a diagnostic. GPQA-Diamond used zero-shot chain-of-thought with flexible-extract. JMMLU reports the common 987-item slice shared by both runs.
B (IFEval): thinking, temperature 1.0, top-p 0.95, top-k 20, min-p 0, presence penalty 1.5, repetition penalty 1.0, maximum 32,768 new tokens, dataset revision 966cd89545d6b6acfd7638bc708b98261ca58e84. The grader reads final-answer content; the saved length includes the complete trace. Length stops were 8 for Qwen3.5-9B and 12 for Qwick-3.5-9B; empty final content occurred 8 and 13 times. This matched comparison replaces all prior temperature-zero IFEval results. It remains report-only, with no numeric selection gate or reselection.
C (LiveCodeBench): temperature 0.6, top-p 0.95, top-k 20, maximum 81,920 new tokens, seed 2026081101. The official runner at commit 28fef95ea8c9f7a547c8329f2cd3d32b92c1fa24 ran in a network-isolated sandbox. There were no generation errors. Length stops were 278 and 76.
D (HMMT): general thinking, maximum 81,920 new tokens, four seeds (2026081101–2026081104), and a fixed boxed/LaTeX/SymPy grader over 60 public problems.
V (complete vision validation): same general-thinking sampling as A, 131,072-token server context, maximum 32,768 new tokens, one to five images per request, and prompt-hash paired seeds. Primary grading is official-compatible final-content parsing, including independent per-lane seed-42 random fallback for MC parse misses; fallback counts and deterministic no-random-fallback strict scores are co-reported.

BF16 complete MMMU validation

This covers every one of the 900 validation rows (30 in each of 30 subjects) from MMMU/MMMU@98e6ac0cb9b7b2cd2c991b85a50762edc4aedc68: 847 multiple-choice and 53 open-answer rows with one to five prompt images. It is the complete validation split, not the larger 10,500-row test split. Official-compatible Qwen3.5-9B/Qwick budget scores were 664/900 and 671/900, with paired transitions 589 both correct, 154 both wrong, 75 Qwen3.5-9B-only, and 82 Qwick-only.

Qwick minus Qwen3.5-9B is +0.778 pp; the subject-stratified paired bootstrap 95% descriptive interval is [-1.889, +3.444] pp and exact two-sided McNemar p=0.6322. Because the interval crosses zero, this run detects no BF16 vision degradation but does not resolve an improvement. The official parser used random fallback on 14 / 21 Qwen3.5-9B/Qwick MC parse misses; deterministic strict budget scores were 662/900 and 667/900. Mean full-trace completion tokens were 6476.48 / 4798.66, a ratio of 0.740937. Qwen3.5-9B/Qwick length stops were 12 / 15. Both lanes have 900/900 raw responses, zero errors, matched seeds, and zero missing full traces. This result remains report-only; it did not trigger model reselection or a holdout rerun.

Public quantized checkpoint vision context

Model	Official-compatible budget	Strict budget	Mean full-trace tokens	Delta vs Qwick BF16
Qwen3.5-9B BF16	664/900 (73.778%)	662/900 (73.556%)	6476.48	-0.778 pp
Qwick BF16	671/900 (74.556%)	667/900 (74.111%)	4798.66	+0.000 pp
Qwick FP8	655/900 (72.778%)	655/900 (72.778%)	5022.40	-1.778 pp
Qwick NVFP4	635/900 (70.556%)	626/900 (69.556%)	5437.84	-4.000 pp

The FP8 and NVFP4 rows use the exported public weights with vision enabled. Their visual towers remain BF16; 200 language-model Linear matrices use the advertised quantization. Qwick BF16/FP8/NVFP4 use the same prompt-hash seed for every item. FP8 minus Qwick BF16 is -1.778 pp with descriptive interval [-4.222, +0.778] pp; the run does not resolve a vision difference. NVFP4 minus Qwick BF16 is -4.000 pp with descriptive interval [-6.778, -1.222] pp; a vision regression signal was detected. All quantized lanes have 900/900 samples, zero errors, and zero missing full traces. These report-only results did not alter selection or trigger a holdout rerun.

Qwen3.5-9B BF16 128k sensitivity diagnostic

Only the 8 upstream BF16 rows that ended with finish_reason=length in the corrected 32k IFEval run were regenerated at a 131,072-token cap. In the logical 541-row overlay, prompt strict changed from 89.279% to 90.018% (+0.739 pp), while length stops and empty final content both changed from 8 to 2. Six rows changed from length to stop; two remained length-limited at 128k.

This is a Qwen3.5-9B-only, report-only sensitivity diagnostic, not a matched Qwick comparison and not a pure causal estimate of the token cap. The generation cap and feasible concurrency/replica layout both changed. All 8 regenerated requests used temperature 1.0, top-p 0.95, top-k 20, min-p 0, presence penalty 1.5, and repetition penalty 1.0; full traces are retained. It did not trigger candidate reselection, and the one-shot synthetic holdout was not rerun.

Frozen internal release gate (separate from public measurement)

The 384-item SFT5 genre-v3 holdout used a separately predeclared internal acceptance policy: thinking, temperature=0, an 8,192-token cap, matched Base/candidate execution, budget accuracy as primary, and fixed accuracy/efficiency thresholds. The candidate weights, data identity, decode, metrics, and thresholds were frozen before generation. The gate passed: Base/candidate budget accuracy was 72.917% / 89.583% and the candidate/Base mean completion-token ratio was 0.484475. These numbers describe only that internal selection condition; they are not estimates under profile A or B and are excluded from the public benchmark table.

The holdout was not used for training or checkpoint selection, no post-result retuning occurred, and it was not rerun. heldout_results.json records all aggregate gate metrics, subgroup results, policy identities, and trace-integrity hashes without redistributing prompts, answers, predictions, or full traces.

Known regressions

HMMT budget correctness fell from 187/240 to 172/240 (-6.25 pp); completion tokens fell by 32.06%.
MMLU-Pro engineering fell from 753/969 to 733/969 (-2.064 pp).
JMMLU common-clean fell from 868/987 to 858/987 (-1.013 pp).

Machine-readable values, decode profiles, and artifact identities are in evaluation_results.json and heldout_results.json.