Qwen3.6-27B Open‑Source: How a 27B Dense Model Outperforms the 397B Giant

Model Overview

Qwen3.6-27B is a 27 B dense multimodal model that surpasses the previous open‑source flagship Qwen3.5‑397B‑A17B on most coding benchmarks.

SWE‑bench Verified: 77.2 (vs 76.2 for 3.5‑397B)

SWE‑bench Pro: 53.5 (vs 50.9)

Terminal‑Bench 2.0: 59.3 (vs 52.5)

SkillsBench Avg5: 48.2 (vs 30.0)

GPQA Diamond: 87.8

AIME 2026: 94.1

Compared with the closed‑source Claude 4.5 Opus, the gap on coding benchmarks is 1‑5 points and Terminal‑Bench scores are identical (59.3).

Key Advantages

Agentic coding: Real‑world coding tasks, especially front‑end and repository‑level, outperform Claude.

Thinking preservation: Multi‑turn conversations keep reasoning context, avoiding repeated “thinking” in iterative coding.

Architecture

Parameters: 27 B dense (no MoE)

Layers: 64, hidden dimension 5120

Native context length: 262 144 tokens, extendable to 1 010 000 tokens

Hidden layout:

16 × (3 × (Gated DeltaNet → FFN) → 1 × (Gated Attention → FFN))

Multimodal vision encoder supports images, video, and documents

Supports MTP (Multi‑Token Prediction) for inference speedup

Gated DeltaNet + Gated Attention mix is more memory‑friendly than pure attention for long contexts.

FP8 Quantized Version

The FP8 weight file is ~30 GB. Using the Qwen/Qwen3.6-27B-FP8 checkpoint halves memory usage while performance loss is reported as negligible.

Why 27 B Is a Sweet Spot

Easy deployment: Dense architecture works directly with vLLM or SGLang without expert parallelism.

Moderate hardware requirements: BF16 needs ~54 GB VRAM (e.g., 2 × A100 40 GB, 1 × H100 80 GB, or 4 × RTX 4090). FP8 needs ~27 GB (single 48 GB L40S/A6000).

No capability compromise: Benchmarks show it outperforms the 397 B model.

Fully open weights: Available on Hugging Face and ModelScope for unrestricted commercial use.

Local Deployment Options

Officially supported routes: vLLM, SGLang, and Hugging Face Transformers. KTransformers also supports CPU‑GPU heterogeneous inference.

vLLM Deployment (recommended)

uv pip install vllm --torch-backend=auto

vllm serve Qwen/Qwen3.6-27B \
  --port 8000 \
  --tensor-parallel-size 8 \
  --max-model-len 262144 \
  --reasoning-parser qwen3

With tool‑call (required for coding agents):

vllm serve Qwen/Qwen3.6-27B \
  --port 8000 \
  --tensor-parallel-size 8 \
  --max-model-len 262144 \
  --reasoning-parser qwen3 \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_coder

Enable MTP (speculative decoding):

vllm serve Qwen/Qwen3.6-27B \
  --port 8000 \
  --tensor-parallel-size 8 \
  --max-model-len 262144 \
  --reasoning-parser qwen3 \
  --speculative-config '{"method":"qwen3_next_mtp","num_speculative_tokens":2}'

Text‑only mode (drops vision encoder):

vllm serve Qwen/Qwen3.6-27B \
  --port 8000 \
  --tensor-parallel-size 8 \
  --max-model-len 262144 \
  --reasoning-parser qwen3 \
  --language-model-only

OOM tip: If out‑of‑memory occurs, do not reduce context below 128 K; the model’s thinking ability degrades sharply.

SGLang Deployment

uv pip install sglang[all]

python -m sglang.launch_server \
  --model-path Qwen/Qwen3.6-27B \
  --port 8000 \
  --tp-size 8 \
  --mem-fraction-static 0.8 \
  --context-length 262144 \
  --reasoning-parser qwen3

With tool use:

python -m sglang.launch_server \
  --model-path Qwen/Qwen3.6-27B \
  --port 8000 \
  --tp-size 8 \
  --mem-fraction-static 0.8 \
  --context-length 262144 \
  --reasoning-parser qwen3 \
  --tool-call-parser qwen3_coder

Enable speculative MTP:

python -m sglang.launch_server \
  --model-path Qwen/Qwen3.6-27B \
  --port 8000 \
  --tp-size 8 \
  --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

Transformers Lightweight Deployment (testing only)

pip install "transformers[serving]"
transformers serve Qwen/Qwen3.6-27B --port 8000 --continuous-batching

This option is suitable for experiments; production should use vLLM or SGLang.

FP8 Quantized Model

Replace the model name with Qwen/Qwen3.6-27B-FP8 and keep the same launch parameters. Example with reduced tensor parallel size:

vllm serve Qwen/Qwen3.6-27B-FP8 \
  --port 8000 \
  --tensor-parallel-size 2 \
  --max-model-len 131072 \
  --reasoning-parser qwen3

Sampling Parameters (official recommendations)

General thinking mode: temperature=1.0, top_p=0.95, top_k=20, presence_penalty=0.0 Precise coding (e.g., WebDev): temperature=0.6, top_p=0.95, top_k=20 Non‑thinking mode:

temperature=0.7, top_p=0.80, top_k=20, presence_penalty=1.5

OpenAI‑Compatible API Usage

from openai import OpenAI
client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY")
messages = [{"role": "user", "content": "Type \"I love Qwen3.6\" backwards"}]
resp = client.chat.completions.create(
    model="Qwen/Qwen3.6-27B",
    messages=messages,
    max_tokens=81920,
    temperature=1.0,
    top_p=0.95,
    presence_penalty=0.0,
    extra_body={"top_k": 20},
)
print(resp)

When thinking mode is enabled, responses include <think>...</think> blocks; switch to non‑thinking parameters to suppress them.

Multimodal Request Example

messages = [{
    "role": "user",
    "content": [
        {"type": "image_url", "image_url": {"url": "https://your-image-url.jpg"}},
        {"type": "text", "text": "这张图里有几个圆？"}
    ]
}]
resp = client.chat.completions.create(
    model="Qwen/Qwen3.6-27B",
    messages=messages,
    max_tokens=81920,
    temperature=1.0,
    top_p=0.95,
    extra_body={"top_k": 20},
)

For video input, replace the type field with video_url.

Pros and Cons

Pros:

27 B dense size enables friendly deployment.

Agentic coding ability surpasses the 397 B MoE model.

Native 262 K context, extendable to 1 M tokens.

Multimodal + text capabilities in a single model.

FP8 quantized version halves memory requirements.

Full‑stack support: vLLM, SGLang, Transformers, KTransformers.

Cons:

Very hard reasoning tasks (e.g., HLE) still favor the 397 B model or Claude 4.5 Opus.

Default thinking mode adds latency; latency‑sensitive production may need to disable it.

Context length should not be reduced below 128 K, otherwise thinking degrades.