Qwen3.5-397B-A17B is Alibaba's Qwen team's follow-up to the Qwen3 generation, released as open weights under an Apache 2.0 license in early 2026. The name states the architecture directly: 397 billion total parameters, with roughly 17 billion active for any given token. It is a sparse Mixture-of-Experts model, and Qwen positions it as a unified vision-language foundation — a single model trained from the start on text, images, and video together, rather than a text model with a vision adapter bolted on afterward.
Under the hood
According to Qwen's own model documentation, the architecture pairs Gated DeltaNet — a linear-attention mechanism — with sparse MoE routing across 60 layers, arranged as repeating blocks of linear-attention-plus-MoE with periodic full-attention layers mixed in. The MoE layer itself carries 512 experts, of which 11 are activated per token: 10 routed experts chosen dynamically plus one shared expert that is always active. That combination — mostly-linear attention for efficiency, MoE for capacity, occasional full attention for the kind of long-range dependency linear attention alone struggles with — is what lets a model with 397 billion total parameters run at roughly the compute cost of a 17-billion-parameter dense model per token.
Context length is 262,144 tokens natively, and Qwen documents extension up to roughly 1,010,000 tokens via YaRN scaling — a technique that adjusts the model's positional encoding to generalize beyond its trained context length, typically with some quality trade-off the further past the native window a request pushes.
Where it shines
Qwen reports strong results on its own benchmark suite: 87.8 on MMLU-Pro, 70.4 on SuperGPQA, and 94.8 on the HMMT February 2025 math competition set, alongside 85.0 on MMMU and 86.7 on MLVU for image and video understanding respectively. These are Qwen's own published figures rather than independently reproduced numbers, so they should be read as the vendor's account of the model's own progress — but taken together they describe a model built for broad, general-purpose reasoning and multimodal understanding rather than a narrow specialist.
The open-weights license is itself a practical advantage: Apache 2.0 permits self-hosting, fine-tuning, and redistribution without the licensing friction that comes with a closed API-only model, which matters for teams that need to run inference on their own infrastructure or adapt the model to a narrow domain.
Where it falls short
Qwen's own documentation frames this release as "cross-generational parity with Qwen3" on many dimensions — meaning the gains over the prior generation are presented as broad and incremental rather than a step-change, and language in the model card is careful not to claim a dramatic leap. As with any Mixture-of-Experts model, the discrepancy between total and active parameters means the model needs far more memory to hold in full than its per-token compute cost would suggest — 397 billion parameters have to be resident somewhere even though only a fraction fire on any given request, which shapes what hardware self-hosting actually requires.
When to pick it
This model fits general-purpose reasoning, coding, and multimodal workloads where a team wants either the option to self-host or the cost profile of a very large model without paying full dense-model compute per request. It is a reasonable default for teams already invested in the Qwen ecosystem, or for workloads that need a genuinely large native context window without leaning on aggressive scaling tricks.
Alternatives worth comparing
The predecessor Qwen3 generation remains available for teams that do not need the added vision and video understanding this release adds. For workloads that need a smaller footprint, Qwen has also shipped considerably smaller models in the same family line that trade some capability for a much lighter memory footprint — worth benchmarking directly against this model on the specific task before committing to the larger one.