Qwen3.5-397B-A17B

Qwen3.5-397B-A17B, published on 16 February 2026, opened the Qwen3.5 generation a week before the rest of the line-up. It is the family's flagship and, unusually for a model of this scale, it is not held back behind an API: the weights are on Hugging Face under Apache 2.0, ungated, alongside an FP8 build that is downloaded three times more often than the full-precision one. The weight files hold 403.4 billion parameters. Each token is routed to ten of 512 experts plus one shared expert, so about 17 billion are active — the same arithmetic as its smaller siblings, applied at a scale that needs a small cluster rather than a workstation. Alibaba benchmarks it against the closed frontier rather than against open models, and the picture is of a model that trades places rather than winning outright. It leads the comparison on IFBench (76.5), MultiChallenge (67.6), NOVA-63 (59.1) and BrowseComp-zh (70.3), and comes close on knowledge and mathematics: MMLU-Pro 87.8 against 89.8 for Gemini 3 Pro, GPQA 88.4 against 92.4 for GPT-5.2, AIME 2026 91.3 against 96.7. It gives clear ground on Humanity's Last Exam (28.7 against 37.5) and on code, where SWE-bench Verified reads 76.4 against 80.9 for Claude 4.5 Opus and LiveCodeBench v6 83.6 against 90.7 for Gemini 3 Pro. On vision it is closer than on text: MMMU 85.0 and MMMU-Pro 79.0, within two points of GPT-5.2 and Gemini 3 Pro on both. All figures come from the vendor. The structure is 60 layers as fifteen repetitions of three Gated DeltaNet blocks and one gated-attention block, each followed by the expert mixture; hidden dimension 4,096, 512 experts of width 1,024, 32 query heads to two key-value heads, vocabulary 248,320, multi-token prediction. Context is 262,144 tokens natively, which the vendor says extends to 1,010,000.

▸Videos

No videos yet.

▸Reviews