Artificial Analysis lists open-weights Qwen3.5 397B A17B
Artificial Analysis has profiled Qwen3.5 397B A17B, a mixture-of-experts model that activates 17B of its 397B parameters per token and ships under Apache 2.0, putting it in the large open-weights class.
Benchmarking firm Artificial Analysis has published a release page for Qwen3.5 397B A17B. The page covers the model's specifications and two variants, one reasoning and one non-reasoning.
The page lists these specifications:
- Parameters: 397B total, 17B active per token
- Context window: 262k tokens
- Input: text and image
- Output: text only
- License: Apache 2.0, with weights on Hugging Face
The one performance figure the page gives is for latency. The non-reasoning variant had the lowest time to first answer token of the two, at 2.44 seconds.
Artificial Analysis compares open-weights models only against others of a similar size. This model falls in its "Large" tier, for models above 150B parameters. The firm's capability indexes draw on evaluations including AA-Omniscience v1.1, AA-Briefcase v1.1, HLE-Diamond, AutomationBench-AA, GDP.pdf, Terminal-Bench 4.0, FrontierCode-AA 2.0 and CritPt v2 (Hard Subset).
The page's text doesn't include the model's scores on any of these indexes. It also doesn't give pricing, output speed or cost per task, though the page has charts for each of these. It doesn't give a release date or say how the reasoning variant compares with the non-reasoning one.
Running 17B of 397B parameters per token keeps inference costs well below what the total parameter count suggests. Together with a permissive license, that makes the model a likely candidate for teams that host their own models. How well it performs against other large open-weights models will depend on the benchmark results, which aren't in the material available.