Summary: Alibaba’s Qwen team released Qwen3.6-35B-A3B as Apache-2.0 open weights on April 16, 2026, pairing 35B total parameters with about 3B active and emphasizing agentic coding plus Thinking Preservation. Self-reported Terminal-Bench 2.0 rose to 51.5 from 40.5 for Qwen3.5-35B-A3B, while SWE-bench Verified reached 73.4. Simon Willison’s laptop comparisons favored Qwen on SVG tasks, but he cautioned that the quantized 21GB model is not generally stronger or more useful than Claude Opus 4.7.
On April 16, 2026, Alibaba’s Qwen team put out Qwen3.6-35B-A3B as open weights — the first open-weight member of the Qwen3.6 series, under the Apache-2.0 license, with 35B total parameters and only about 3B active for any given token. The announcement’s title is refreshingly blunt: “Agentic coding power, now open to all.” It marks a second “large total, small active” delivery since February’s Qwen3.5, and this time agentic coding is the selling point. Anthropic shipped Claude Opus 4.7 that same day, and Simon Willison promptly raced the two on a laptop — a head-to-head that had plenty of observers repricing open weights on the spot.
Qwen3.6’s Debut Open Weights
Per the model card, community feedback directly shaped Qwen3.6, which puts stability and real-world utility first. Two upgrades lead the changelog. One is agentic coding: frontend workflows and repository-level reasoning now flow with less friction. The other is Thinking Preservation, a mechanism that keeps the reasoning context carried in historical messages, so iterative development no longer re-derives its thinking every round. The model is multimodal — a vision encoder takes images and video as input — and ships with a native context window of 262,144 tokens, stretchable to 1,010,000.
Linear Attention Meets 256 Experts
Inside, it is a 40-layer MoE with 256 experts, of which each token activates 8 routed experts plus 1 shared one. The arrangement blends two attention styles: every macro block stacks three “Gated DeltaNet into MoE” stages before a single “Gated Attention into MoE” stage, and the pattern repeats ten times. Linear attention holds down the compute cost of long contexts, while sparse GQA keeps the retrieval sharpness that matters. That pairing is what lets a million-token context and laptop-runnable share one spec sheet. A 248,320-token embedding vocabulary rounds out the multilingual, multimodal positioning.
Benchmark Scores for Agentic Coding
The headline gain shows up in Terminal-Bench 2.0: 51.5, well clear 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 comes in at 73.4 — just under Qwen3.5-27B’s 75.0, but clear of the earlier A3B’s 70.0 — alongside SWE-bench Multilingual at 67.2 and SWE-bench Pro at 49.5. QwenWebBench, the in-house front-end code-generation suite, yields a 1397 rating, and the model’s MCPMark 37.0 tops its comparison set. Math and knowledge stay solid: AIME26 92.7, GPQA 86.0, LiveCodeBench v6 80.4. On vision, MMMU 81.7 and OmniDocBench1.5 89.9 — document understanding that agents reading screenshots and files will put straight to work.
The Pelican Benchmark, Run at Home
Later that afternoon, Simon Willison loaded Unsloth’s 20.9GB Q4_K_S GGUF quantization onto his MacBook Pro M5 through LM Studio and entered it against the just-shipped Claude Opus 4.7 in his pelican-riding-a-bicycle SVG benchmark. Qwen took the round — Opus botched the bicycle frame — and even a rematch with thinking_level set to max didn’t rescue it. Because labs might have trained on his joke test, he pulled out a confidential spare, a flamingo riding a unicycle; Qwen won that one too, tucking an SVG comment reading “Sunglasses on flamingo!” in for good measure. Willison’s read stays level-headed: he strongly doubts a 21GB quantized model offers more power or usefulness 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 collected 463 points on Hacker News; the release announcement itself climbed to 1,274 points and 532 comments.
Running It, and the Recommended Settings
The weights slot straight into Transformers, vLLM, SGLang, and KTransformers; bring in Qwen-Agent for agentic workflows, or hook up any OpenAI-compatible endpoint. The model card’s suggested settings: temperature 1.0 for thinking mode as a default, dropping to 0.6 for precise coding tasks, with top_p at 0.95; plan on 32,768 output tokens for typical queries and 81,920 when the problem is competition-grade. For the local-first crowd, the punchline is a 20.9GB file driving a near-frontier coding agent on an ordinary consumer laptop — exactly the trajectory the open-source split shaping 2026 sketched at the start of the year: open weights keep ferrying frontier capability onto local hardware, while the ground where closed models hold a clear edge narrows toward inference at massive scale.