Alibaba Qwen Releases Qwen3.6-27B Outperforming Its Previous 397B Flagship in Coding

QwenQwen

· Updated

Alibaba open-sourced Qwen3.6-27B, a dense multimodal model that surpasses the previous 397B-parameter flagship on every major agentic coding benchmark. The release proves that optimized mid-sized dense models can deliver frontier-level performance while remaining significantly easier to deploy than massive mixture-of-experts architectures.

Alibaba released Qwen3.6-27B under an Apache 2.0 license, a dense model (where every parameter is active for every token) featuring native multimodality and reasoning. It follows the launch of Qwen3.6-Plus and the sparse Qwen3.6-35B-A3B as part of the team's latest model family.
Parameters
27B (Dense)
Context window
200K tokens
License
Apache 2.0
Benchmark (SWE-bench Verified)
77.2%
Benchmark (Terminal-Bench 2.0)
59.3%
Benchmark (GPQA Diamond)
87.8%
Benchmark (AIME 26)
94.1%

The 27B model outperforms the previous generation's 397B-parameter MoE flagship on coding benchmarks. Delivering flagship-level agentic coding in a smaller footprint removes the deployment complexity associated with massive mixture-of-experts (architectures that only activate a fraction of parameters per token).

You can access the model via Qwen Studio, Alibaba Cloud's API, or by downloading weights from Hugging Face, matching the local execution support in Ollama. It includes a new preserve_thinking feature that maintains reasoning traces across multi-turn interactions, making it a drop-in backend for agents like Claude Code.

View the full update on qwen.ai
Qwen
Qwen
@Alibaba_Qwen
X

🚀 Meet Qwen3.6-27B, our latest dense, open-source model, packing flagship-level coding power! Yes, 27B, and Qwen3.6-27B punches way above its weight. 👇 What's new: 🧠 Outstanding agentic coding — surpasses Qwen3.5-397B-A17B across all major coding benchmarks 💡 Strong reasoning across text & multimodal tasks 🔄 Supports thinking & non-thinking modes ✅ Apache 2.0 — fully open, fully yours Smaller model. Bigger results. Community's favorite. ❤️ We can't wait to see what you build with Qwen3.6-27B! 👀 🔗👇 Blog: https://t.co/P2Zx7FwMxB Qwen Studio: https://t.co/c4vm4LuZrU Github: https://t.co/zKDEbv0R4U Hugging Face: https://t.co/N67hyzxvfr https://t.co/SSdtbWRDap ModelScope: https://t.co/xODf1pj9kw https://t.co/xXhoqlJ2AB

1.6kretweets12klikes
View on X

Still wondering? A few quick answers below.

Qwen3.6-27B is a dense, 27-billion-parameter multimodal model released by Alibaba. Unlike mixture-of-experts models that only activate some parameters, this dense architecture is straightforward to deploy. It supports text, image, and video inputs and features an internal reasoning process called thinking mode to handle complex logic and agentic coding tasks with high accuracy.

Despite being roughly 15 times smaller in total parameter count, Qwen3.6-27B outperforms the previous 397-billion-parameter flagship across all major agentic coding benchmarks. This includes higher scores on SWE-bench Verified and Terminal-Bench 2.0. Its dense architecture makes it easier to serve and fine-tune compared to the massive mixture-of-experts structure of the older generation.

Yes, Qwen3.6-27B is released under the Apache 2.0 license, making it fully open for community use and commercial applications. The model weights are available for download on Hugging Face and ModelScope. This permissive licensing allows developers to self-host the model on their own infrastructure or integrate it into private development environments without proprietary restrictions.

The preserve_thinking feature allows the model to maintain its internal reasoning traces from previous conversation turns during multi-step interactions. This is specifically recommended for agentic tasks where the model must remember its logic and planning steps across a long dialogue. By keeping these thinking tokens in context, the model achieves more reliable performance in complex autonomous workflows.

Qwen3.6-27B is designed for deep integration with popular agentic coding assistants. It is natively compatible with OpenClaw, Qwen Code, and Claude Code via the Anthropic-compatible API protocol. Developers can use it as a backend for terminal-based agents that autonomously navigate codebases, execute commands, and edit files across multiple steps to solve complex software engineering issues.

Every HeadsUpAI update is written based on its original source and reviewed before it's published. Read our editorial standards →

See all AI news & updates from Qwen →

Share this update