返回全部资讯
Qwen 官方博客 (通义)T1·基座大模型

Qwen3.6-27B 开源:稠密模型编程能力超越前代旗舰

作者:QwenTeam原标题:Qwen3.6-27B:270亿参数稠密模型,旗舰级编程能力
85AI 研判分

通义千问团队正式开源 Qwen3.6-27B,这是一款拥有 270 亿参数的稠密多模态模型。该模型在智能体编程任务上达到旗舰级表现,全面超越前代 MoE 架构的 Qwen3.5-397B-A17B。作为稠密架构,它无需复杂的 MoE 路由即可部署,为开发者提供了在实用规模下获取顶尖编程能力的理想选择。

报道正文

本文目录12 个章节

Qwen3.6-27B Main Image

QWEN STUDIO HUGGING FACE MODELSCOPE DISCORD

继 Qwen3.6-Plus 和 Qwen3.6-35B-A3B 发布之后,我们非常高兴地宣布开源 Qwen3.6-27B —— 一个拥有270亿参数的稠密多模态模型,也是社区呼声最高的模型规格。Qwen3.6-27B 依然支持多模态思考与非思考模式,在智能体编程方面达到了旗舰级表现,全面超越前代开源旗舰 Qwen3.5-397B-A17B(总参数397B / 激活参数17B的MoE模型)。作为稠密架构,它无需MoE路由即可部署,是开发者在实用、可广泛部署规模上获取顶尖编程能力的理想选择。Qwen3.6-27B 现已在 Qwen Studio 上线,可通过 API 调用,并以开源权重的形式向社区发布。

模型表现#

下文将全面展示 Qwen3.6-27B 与其他稠密及 MoE 基线模型(包括前代开源旗舰 Qwen3.5-397B-A17B)的评测对比结果。Qwen3.6-27B 在智能体编程基准上取得了显著突破,超越了总参数量高达其15倍的模型。

自然语言#

Qwen3.6-27B 在稠密模型的智能体编程能力上实现了突破。仅凭270亿参数,它在所有主要编程基准上全面超越了 Qwen3.5-397B-A17B(总参数397B / 激活参数17B)——包括 SWE-bench Verified(77.2 vs. 76.2)、SWE-bench Pro(53.5 vs. 50.9)、Terminal-Bench 2.0(59.3 vs. 52.5)以及 SkillsBench(48.2 vs. 30.0)。同时,它也大幅领先于同规模的稠密模型。在推理任务上,Qwen3.6-27B 在 GPQA Diamond 上取得了87.8的成绩,可与数倍于其规模的模型相媲美。

Qwen3.5-27BQwen3.5-397B-A17BGemma4-31BClaude 4.5 OpusQwen3.6-35B-A3BQwen3.6-27B
Coding Agent
SWE-bench Verified75.076.252.080.973.477.2
SWE-bench Pro51.250.935.757.149.553.5
SWE-bench Multilingual69.369.351.777.567.271.3
Terminal-Bench 2.041.652.542.959.351.559.3
SkillsBench Avg527.230.023.645.328.748.2
QwenWebBench106811861197153613971487
NL2Repo27.332.215.543.229.436.2
Claw-Eval Avg64.370.748.576.668.772.4
Claw-Eval Pass^346.248.125.059.650.060.6
QwenClawBench52.251.841.752.352.653.4
Knowledge
MMLU-Pro86.187.885.289.585.286.2
MMLU-Redux93.294.993.795.693.393.5
SuperGPQA65.670.465.770.664.766.0
C-Eval90.593.082.692.290.091.4
STEM & Reasoning
GPQA Diamond85.588.484.387.086.087.8
HLE24.328.719.530.821.424.0
LiveCodeBench v680.783.680.084.880.483.9
HMMT Feb 2592.094.888.792.990.793.8
HMMT Nov 2589.892.787.593.389.190.7
HMMT Feb 2684.387.977.285.383.684.3
IMOAnswerBench79.980.974.584.078.980.8
AIME2692.693.389.295.192.794.1
  • SWE-Bench Series: Internal agent scaffold (bash + file-edit tools); temp=1.0, top_p=0.95, 200K context window. We correct some problematic tasks in the public set of SWE-bench Pro and evaluate all baselines on the refined benchmark.
  • Terminal-Bench 2.0: Harbor/Terminus-2 harness; 3h timeout, 32 CPU/48 GB RAM; temp=1.0, top_p=0.95, top_k=20, max_tokens=80K, 256K ctx; avg of 5 runs.
  • SkillsBench: Evaluated via OpenCode on 78 tasks (self-contained subset, excluding API-dependent tasks); avg of 5 runs.
  • NL2Repo: Others are evaluated via Claude Code (temp=1.0, top_p=0.95, max_turns=900).
  • QwenClawBench: A real-user-distribution Claw agent benchmark; temp=0.6, 256K ctx.
  • QwenWebBench: An internal front-end code generation benchmark; bilingual (EN/CN), 7 categories (Web Design, Web Apps, Games, SVG, Data Visualization, Animation, and 3D); auto-render + multimodal judge (code/visual correctness); BT/Elo rating system.
  • AIME 26: We use the full AIME 2026 (I & II), where the scores may differ from Qwen 3.5 notes.

视觉语言#

Qwen3.6-27B 原生支持多模态,支持视觉语言思考与非思考模式——与 Qwen3.6-35B-A3B 相同。它能够处理图像、视频与文本的多模态理解,支持视觉推理、文档理解和视觉问答等任务。

Qwen3.5-27BQwen3.5-397B-A17BGemma4-31BClaude 4.5 OpusQwen3.6-35B-A3BQwen3.6-27B
STEM & Puzzle
MMMU82.385.080.480.781.782.9
MMMU-Pro75.079.076.970.675.375.8
MathVista mini87.8--79.3--86.487.4
DynaMath87.786.379.579.782.885.6
VlmsAreBlind96.9--87.2--96.697.0
General VQA
RealWorldQA83.783.972.377.085.384.1
MMStar81.083.877.373.280.781.4
MMBenchEN-DEV-v1.192.6--90.9--92.892.3
SimpleVQA56.067.152.965.758.956.1
Document Understanding
CharXiv RQ79.580.867.968.578.078.4
CC-OCR81.082.075.776.981.981.2
OCRBench89.4--86.1--90.089.4
Spatial Intelligence
ERQA60.567.557.546.861.862.5
CountBench97.897.296.190.696.197.8
RefCOCO avg90.992.3----92.092.5
EmbSpatialBench84.5------84.384.6
RefSpatialBench67.7--4.7--64.370.0
Video Understanding
VideoMME(w sub.)87.087.5--77.786.687.7
VideoMMMU82.384.781.684.483.784.4
MLVU85.986.7--81.786.286.6
MVBench74.677.6--67.274.675.5
Visual Agent
V*93.795.8--67.090.194.7
AndroidWorld64.2--------70.3
  • Empty cells (--) indicate scores not yet available or not applicable.

开始使用 Qwen3.6-27B#

Qwen3.6-27B 的开源权重已在 Hugging Face 和 ModelScope 上提供,支持本地部署;也可通过 阿里云百炼 API 调用。此外,您还可以在 Qwen Studio 上即时体验。

该模型可以无缝集成到流行的第三方编程助手中,包括 OpenClaw、Claude Code 和 Qwen Code,从而简化开发流程,实现高效且具备上下文感知能力的编码体验。

API 使用方式#

本次发布支持 preserve_thinking 功能:在消息中保留所有前序轮次的思维内容,推荐用于智能体任务。

阿里云百炼#

阿里云百炼支持行业标准协议,包括兼容 OpenAI 规范的聊天补全(chat completions)和响应(responses)API,以及兼容 Anthropic 的 API 接口。

以下是聊天补全 API 的代码示例:

PYTHON
"""
Environment variables (per official docs):
  DASHSCOPE_API_KEY: Your API Key from https://bailian.console.aliyun.com/
  DASHSCOPE_BASE_URL: (optional) Base URL for compatible-mode API.
    - Beijing: https://dashscope.aliyuncs.com/compatible-mode/v1
    - Singapore: https://dashscope-intl.aliyuncs.com/compatible-mode/v1
    - US (Virginia): https://dashscope-us.aliyuncs.com/compatible-mode/v1
  DASHSCOPE_MODEL: (optional) Model name; override for different models.
"""
from openai import OpenAI
import os

api_key = os.environ.get("DASHSCOPE_API_KEY")
if not api_key:
    raise ValueError(
        "DASHSCOPE_API_KEY is required. "
        "Set it via: export DASHSCOPE_API_KEY='your-api-key'"
    )

client = OpenAI(
    api_key=api_key,
    base_url=os.environ.get(
        "DASHSCOPE_BASE_URL",
        "https://dashscope.aliyuncs.com/compatible-mode/v1",
    ),
)

messages = [{"role": "user", "content": "Introduce vibe coding."}]

model = os.environ.get(
    "DASHSCOPE_MODEL",
    "qwen3.6-27b",
)
completion = client.chat.completions.create(
    model=model,
    messages=messages,
    extra_body={
        "enable_thinking": True,
        # "preserve_thinking": True,
    },
    stream=True
)

reasoning_content = ""  # Full reasoning trace
answer_content = ""  # Full response
is_answering = False  # Whether we have entered the answer phase
print("\n" + "=" * 20 + "Reasoning" + "=" * 20 + "\n")

for chunk in completion:
    if not chunk.choices:
        print("\nUsage:")
        print(chunk.usage)
        continue

delta = chunk.choices[0].delta

# Collect reasoning content only
    if hasattr(delta, "reasoning_content") and delta.reasoning_content is not None:
        if not is_answering:
            print(delta.reasoning_content, end="", flush=True)
        reasoning_content += delta.reasoning_content

# Received content, start answer phase
    if hasattr(delta, "content") and delta.content:
        if not is_answering:
            print("\n" + "=" * 20 + "Answer" + "=" * 20 + "\n")
            is_answering = True
        print(delta.content, end="", flush=True)
        answer_content += delta.content

更多信息请访问 API 文档。

代码及智能体#

Qwen3.6-27B 具备出色的智能体编程能力,可以无缝集成到流行的第三方编程助手中,包括 OpenClaw、Claude Code 和 Qwen Code。

OpenClaw#

Qwen3.6-27B 兼容 OpenClaw(原名 Moltbot / Clawdbot),这是一款可自托管的开源 AI 编码智能体。将其连接至 百炼,即可在终端中获得完整的智能体编码体验。请使用以下脚本开始:

BASH
# Node.js 22+
curl -fsSL https://molt.bot/install.sh | bash   # macOS / Linux

# Set your API key
export DASHSCOPE_API_KEY=<your_api_key>

# Launch OpenClaw
openclaw dashboard # web browser
# openclaw tui # Open a new terminal and start the TUI

首次使用时,请编辑 ~/.openclaw/openclaw.json 文件,将 OpenClaw 指向百炼。 找到或创建以下字段并合并它们——切勿覆盖整个文件,以保留您现有的设置:

JSON
{
  "models": {
    "mode": "merge",
    "providers": {
      "bailian": {
        "baseUrl": "https://dashscope.aliyuncs.com/compatible-mode/v1",
        "apiKey": "DASHSCOPE_API_KEY",
        "api": "openai-completions",
        "models": [
          {
            "id": "qwen3.6-27b",
            "name": "qwen3.6-27b",
            "reasoning": true,
            "input": ["text", "image"],
            "contextWindow": 131072,
            "maxTokens": 16384
          }
        ]
      }
    }
  },
  "agents": {
    "defaults": {
      "model": {
        "primary": "bailian/qwen3.6-27b"
      },
      "models": {
        "bailian/qwen3.6-27b": {}
      }
    }
  }
}

Qwen Code#

Qwen3.6-27B 适配 Qwen Code,这是一款专为终端设计的开源 AI 智能体,针对 Qwen 系列进行了深度优化。请使用以下脚本开始:

BASH
# Node.js 20+
npm install -g @qwen-code/qwen-code@latest

# Start Qwen Code (interactive)
qwen

# Then, in the session:
/help
/auth

首次使用时,系统会提示您登录。您可以随时运行 /auth 来切换认证方式。

Claude Code#

Qwen API 也支持 Anthropic API 协议,这意味着您可以将其与 Claude Code 等工具配合使用,以获得更优质的编码体验:

BASH
# Install Claude Code
npm install -g @anthropic-ai/claude-code

# Configure environment
export ANTHROPIC_MODEL="qwen3.6-27b"
export ANTHROPIC_SMALL_FAST_MODEL="qwen3.6-27b"
export ANTHROPIC_BASE_URL=https://dashscope.aliyuncs.com/apps/anthropic
export ANTHROPIC_AUTH_TOKEN=<your_api_key>

# Launch the CLI
claude

总结#

Qwen3.6-27B 的发布,证明了一个经过精心训练的稠密模型,可以在开发者最关心的任务上,超越规模显著更大的前代模型。作为广泛部署的开源模型规格——270亿参数,它在所有主要智能体编程基准上超越了拥有3970亿参数的 Qwen3.5-397B-A17B,同时部署和服务都更加便捷。随着 Qwen3.6-27B 正式加入,Qwen3.6 开源家族现已构建起覆盖全尺度的模型矩阵,这也凸显了智能体编程技术全面跃升的一代——从30亿激活参数的 Qwen3.6-35B-A3B,到线上的 Qwen3.6-Plus 和 Qwen3.6-Max-Preview,这一代在各个规模上都实现了智能体编程能力的飞跃。我们由衷感谢社区的宝贵反馈,并期待看到大家利用这些模型创造出的精彩成果。敬请关注 Qwen 团队的后续发布!

引用#

如果 Qwen3.6-27B 对你有所帮助,欢迎引用以下文章:

BIBTEX
@misc{qwen36_27b,
    title = {{Qwen3.6-27B}: Flagship-Level Coding in a 27B Dense Model},
    url = {https://qwen.ai/blog?id=qwen3.6-27b},
    author = {{Qwen Team}},
    month = {April},
    year = {2026}
}