On April 22, the Qwen team released a new model on Hugging Face. It has 27B parameters, a Dense architecture, and an Apache 2.0 license. It is called Qwen3.6-27B.
In the release blog post, they listed benchmark scores. SWE-bench Verified: 77.2. For comparison, they also included the score for Qwen3.5-397B-A17B: 76.2.
A 27B Dense model beat its own 397B MoE flagship on a coding benchmark.
This is not a coincidence. There are specific reasons behind it.
Where MoE Starts to Leak
MoE (Mixture of Experts) architecture has been the mainstream solution for scaling large models over the past two years. The logic is simple: the total parameter count is large, but only a small subset of "experts" is activated per inference, keeping the actual computation relatively manageable.
Qwen3.5-397B has 397B total parameters and activates 17B per inference. Qwen3.6-27B activates all 27B per inference—the computational cost is similar, but the approach is fundamentally different.
Inference with a Dense architecture is more consistent and predictable. MoE has some randomness in routing: different inputs may trigger different combinations of experts, which can accumulate deviations in long-chain tasks.
Stability is critical for tasks like code generation. You want "correct at every step," not "good on average but sometimes drifts."
On Terminal-Bench 2.0, the 27B model scored 59.3, while the 397B model scored only 52.5. The gap was even larger on SkillsBench: 48.2 versus 30.0.
| Model | Parameter Scale | SWE-bench | Terminal-Bench | SkillsBench |
|---|---|---|---|---|
| Qwen3.6-27B | 27B Dense | 77.2 | 59.3 | 48.2 |
| Qwen3.5-397B-A17B | 397B MoE | 76.2 | 52.5 | 30.0 |
The 27B model won all three tests.
What is Thinking Preservation
Qwen3.6-27B includes a mechanism called "Thinking Preservation" that deserves a separate mention.
In multi-turn conversations, the model retains a summary of its previous reasoning process. In the next turn, it does not need to reason from scratch—it continues directly from the previous thinking results.
This is very useful in agent workflows. If an agent needs to complete a ten-step task, the analysis from the first three steps should not disappear by step four; it should persist as context. This reduces repeated token generation and lowers "forgetting drift" in long tasks.
This design moves in the opposite direction from MoE's routing mechanism: MoE scales parameters horizontally, while Thinking Preservation optimizes state transfer vertically. The two approaches solve problems at different levels.
Runs on Consumer Hardware
This is another practical value of the model.
The full version of Qwen3.5-397B requires 807 GB of storage. Even a quantized version needs hundreds of GB, requiring a multi-GPU server to run, which is out of reach for most individual developers.
Qwen3.6-27B:
- Full version: 55.6 GB
- Quantized GGUF version: 16.8 GB
- Runs on a single RTX 4090 at about 25 tokens/second
- Works on an M4 Max MacBook without issue
The default context window is 262,144 tokens, expandable to 1 million. The Apache 2.0 license allows commercial use without restrictions.
This significantly lowers the barrier for local deployment. Engineering teams that want to deploy a coding agent on their intranet without sending code to a cloud API previously had to choose between poor open-source models or expensive GPU servers. Now there is an option.
What This Means for MoE
The timing of Alibaba's release is interesting. Qwen3.6-27B followed a few days after Qwen3.6-Max (the flagship closed-source version)—one showcases top-line capability to attract commercial clients, the other serves the developer community. Two legs, serving different audiences.
But the bigger question is: If a 27B Dense model can already go toe-to-toe with Claude Opus 4.5 on agentic coding, on what dimension will the next competition be fought?
On Terminal-Bench, the 27B model tied with Claude Opus 4.5, while Opus 4.5's API costs over $15 per million tokens. Qwen3.6-27B is open-source and can be self-deployed.
Incremental gains on benchmarks are shrinking. Users are increasingly concerned with latency, cost, and the feasibility of local deployment.
This trend favors open-source models and puts pressure on closed-source API services.
The next step is to see how Kimi, DeepSeek, and MiniMax respond.
Sources: CocoLoop, Qwen3.6-27B: Flagship-Level Coding in a 27B Dense Model (Simon Willison's Blog); Alibaba's Qwen3.6-27B Beats 397B Model in Coding Tasks (AI Daily Post); Qwen3.6-27B (Hacker News)