Let's start with the number: SWE-bench Pro 53.5 vs 50.9.
The former is Qwen3.6-27B, a 27-billion-parameter model open-sourced on April 22 under the Apache 2.0 license. The latter is Qwen3.5-397B, a 397-billion-parameter Mixture-of-Experts (MoE) model that was previously Alibaba's flagship.
A small model with 27 billion parameters beating a large model with nearly 40 billion activated parameters on a software engineering benchmark is no small feat.
Thinking Preservation: The Key Innovation
To be honest, the story of a smaller model outperforming a larger one has been told several times in this field. What makes Qwen3.6-27B's achievement stand out is not the architecture itself, but a new feature: Thinking Preservation.
There is an overlooked problem in open-source LLMs: when an agent runs multi-turn tasks, the reasoning process from each turn is discarded. The model has to re-infer from context each time, rather than reusing previously established cognitive chains.
Qwen3.6-27B's approach is to retain the "thinking traces" from historical messages as persistent context:
"retains and leverages thinking traces from historical messages across the entire conversation rather than discarding prior reasoning, improving efficiency in multi-turn agent workflows"
In plain language: the model remembers how it thought before, instead of resetting to zero each time.
On agentic coding tasks, this difference is significant. When working on the same code repository, the gap between a model that remembers its previous reasoning conclusions and one that does not is large. Since SWE-bench Pro tests real codebase tasks, this advantage is amplified.
Benchmark Overview
Qwen3.6-27B's performance is not limited to SWE-bench:
| Benchmark | Qwen3.6-27B | Comparison |
|---|---|---|
| SWE-bench Pro | 53.5 | vs Qwen3.5-397B MoE: 50.9 |
| Terminal-Bench 2.0 | 59.3 | On par with Claude Opus flagship |
| AIME26 (Math Competition) | 94.1 | vs Qwen3.5-27B: 92.6 |
| QwenWebBench | 1487 | vs Qwen3.5-27B: 1068 (+39%) |
| SkillsBench Average | 48.2 | 77% improvement over previous generation same-size model |
| GPQA Diamond | 87.8 | vs Qwen3.5-27B: 85.5 |
The 77% improvement on SkillsBench deserves special mention. This benchmark measures cross-domain general capability. The jump from Qwen3.5-27B to Qwen3.6-27B is not a minor tweak; it is a system-level reconstruction.
Context Window
Native support for 262,144 tokens (approximately 200,000 Chinese characters), scalable to 1,010,000 tokens using YaRN.
For scenarios requiring long code repositories or long documents, this context is sufficient.
What Open Source Means
Apache 2.0 license, unrestricted commercial use.
This is more practical for enterprises than performance numbers alone: a 27-billion-parameter model costs far less to deploy locally than a 397-billion-parameter model, yet scores higher on core tasks — a combination that is attractive for companies looking to build their own AI capabilities. Consider that a single consumer-grade GPU can run 27B, while 397B requires at least multiple high-end machines.
This is also why the 77% SkillsBench number is crucial: it is not just stronger on code, but general capabilities have comprehensively improved. That is the real prerequisite for replacing a large model.
This Week's Landscape
This week, OpenAI released GPT-5.5 and Alibaba released Qwen3.6-27B.
Looking at both together: closed-source flagships are pursuing "faster and more token-efficient," while open-source small models are pursuing "better results with fewer parameters." Different directions, but both are squeezing the other's space.
The next thing to be disrupted may not be a specific company, but the assumption that "you need a large model to run code well."
Sources: CocoLoop, Alibaba Qwen Team Releases Qwen3.6-27B: A Dense Open-Weight Model Outperforming 397B MoE on Agentic Coding Benchmarks (MarkTechPost)