On April 16, 2026, Alibaba’s Qwen team open-sourced Qwen3.6-35B-A3B — the first open-weight variant of the Qwen3.6 series, released under Apache-2.0, with 35B total parameters and only about 3B activated per token. The blog title does not hedge: “Agentic coding power, now open to all.” It is the latest “big total, small active” drop since February’s Qwen3.5 series, but the headline use case has shifted to agentic coding. Anthropic released Claude Opus 4.7 the same morning, and a same-day laptop showdown between the two, run by Simon Willison, made a lot of people reprice where open weights now sit.
The First Open-Weight Qwen3.6
According to the model card, Qwen3.6 was built on direct community feedback and prioritizes stability and real-world utility. Two upgrades get top billing. First, agentic coding: frontend workflows and repository-level reasoning are handled with greater fluency. Second, Thinking Preservation — a new option to retain reasoning context from historical messages, so iterative development stops recomputing its own thinking and the overhead that comes with it. The model is multimodal, pairing the language model with a vision encoder that accepts images and video, with a native context of 262,144 tokens, extensible to 1,010,000.
A Hybrid of Linear Attention and 256 Experts
Under the hood it is a 40-layer MoE: 256 experts, with 8 routed plus 1 shared expert active per token. The layout mixes attention styles — each macro block is three “Gated DeltaNet into MoE” stages followed by one “Gated Attention into MoE” stage, repeated ten times. Linear attention keeps long-context compute cheap; sparse GQA attention preserves the retrieval precision that matters. That combination is why a million-token context and “runs on a laptop” can appear on the same spec sheet. The 248,320-token embedding vocabulary signals the multilingual, multimodal positioning.
Agentic Coding Benchmark Scores
The sharpest jump is Terminal-Bench 2.0: 51.5, clearly ahead of Qwen3.5-35B-A3B at 40.5 and Qwen3.5-27B at 41.6 (Harbor/Terminus-2 harness, averaged over five runs). SWE-bench Verified lands at 73.4 — slightly below Qwen3.5-27B’s 75.0 but above the previous A3B’s 70.0 — with SWE-bench Multilingual at 67.2 and SWE-bench Pro at 49.5. On QwenWebBench, the internal front-end code-generation suite, it scores a 1397 rating, and its MCPMark 37.0 is the best in its comparison set. Math and knowledge hold: AIME26 92.7, GPQA 86.0, LiveCodeBench v6 80.4. On the vision side, MMMU 81.7 and OmniDocBench1.5 89.9 — document understanding that agents reading screenshots and files will notice.
The Pelican Test, on a Laptop
That same afternoon, Simon Willison ran the model as a 20.9GB Q4_K_S GGUF quantization by Unsloth on his MacBook Pro M5 through LM Studio, and pitted it against the freshly released Claude Opus 4.7 on his pelican-riding-a-bicycle SVG benchmark. Qwen won — Opus mangled the bicycle frame — and a second Opus attempt with thinking_level set to max did not do much better. Suspecting labs might train for his joke benchmark, he burned a secret backup test, a flamingo riding a unicycle; Qwen won again, earning extra credit for an SVG comment reading “Sunglasses on flamingo!” Willison stays sober about it: he very much doubts a 21GB quantized model is more powerful or useful than Anthropic’s latest proprietary release, but “if the thing you need is an SVG illustration of a pelican riding a bicycle, right now Qwen3.6-35B-A3B running on a laptop is a better bet than Opus 4.7.” His post drew 463 points on Hacker News; the release announcement itself hit 1,274 points and 532 comments.
Deployment and Best Practices
The weights work with Transformers, vLLM, SGLang, and KTransformers; agentic use pairs with Qwen-Agent, and any OpenAI-compatible endpoint will do. Recommended settings from the model card: temperature 1.0 for thinking mode in general, 0.6 for precise coding tasks, top_p 0.95; output length of 32,768 tokens for most queries, stretched to 81,920 for competition-grade problems. For local-first developers, the takeaway is a 20.9GB file running a near-frontier coding agent on consumer laptop hardware — consistent with how the open-source split shaping 2026 looked from the start of the year: open weights keep pushing frontier capability onto local hardware, and the territory where closed models are clearly ahead keeps shrinking toward tasks that need massive-scale inference.
Sources
- Qwen3.6-35B-A3B: Agentic coding power, now open to all — Qwen
- Qwen/Qwen3.6-35B-A3B — Hugging Face
- Qwen3.6-35B-A3B on my laptop drew me a better pelican than Claude Opus 4.7 — Simon Willison
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.