Alibaba Releases Open-Source Qwen3.6-27B Coding Model
Alibaba has released Qwen3.6-27B, a 27 billion-parameter dense open-source model under an Apache 2.0 license. The company said the model surpasses Qwen3.5-397B-A17B on major agentic coding benchmarks including SWE-bench and Terminal-Bench 2.0, while also processing text, images and video and offering both "thinking" and "non-thinking" modes from a single checkpoint.
Early performance reports suggested the model can run on relatively modest hardware. One version was reported to fit on a 16GB system with a 32,000-token context window, and a separate baseline test using standard settings measured 51.83 tokens per second on an Nvidia RTX 5090, 45.59 on an RTX 4090, 22.30 on Apple's M2 Ultra and 11.08 on a DGX Spark.
From the sources (4 posts)
@himanshustwtsRT @atomic_chat_hq: Compared Qwen3.6 35B and 27B in the same conditions with Google TurboQuant Device: MacBook Pro M5Max 64GB RAM Outputs…
@clementdelangueRT @coffeecup2020: Qwen3.6-27B-TQ3_4S is insanely good! fit on my 16GB with 32k context Two prompts and I get th…
@stevibeQwen3.6 27B landed yesterday, so I ran it on 4 setups side-by-side to see how they stack up: 🔴 RTX 4090 — 45.59 tok/s, TTFT 525ms 🟢 RTX 5090 — 51.83 tok/s, TTFT 752ms ⚫️ M2 Ultra — 22.30 tok/s, TTFT 216ms 🟣 DGX Spark — 11.08 tok/s, TTFT 31
@wesrothAlibaba launched Qwen3.6-27B, a highly efficient, dense open-source model (licensed under Apache 2.0) packing 27 billion parameters. Despite its relatively small footprint, it decisively outperforms its massive predecessor across every maj