Alibaba's new 397B open-weight multimodal monster

Qwen3.5-397B-A17B is the first open-weight model in Alibaba's Qwen3.5 lineup, built natively multimodal and specifically tuned for real-world agent workflows. Its main architectural feat is efficiency: while holding 397 billion total parameters, its sparse Mixture-of-Experts setup activates just 17 billion per token.

Big model, tiny active footprint

To balance massive model capacity with viable inference costs, the model pairs hybrid linear attention with sparse MoE, bolstered by large-scale reinforcement learning environment scaling.

According to the Qwen team, this architecture boosts decoding throughput by 8.6x to 19.0x compared to Qwen3-Max, while extending language coverage to 201 languages and dialects.

The weights and code are released under a permissive Apache 2.0 license. Everything is available to download on GitHub, Hugging Face, and ModelScope, and it can also be tested hosted through Qwen Chat and their API.

Why it matters

For teams building and self-hosting autonomous multimodal agents, this is an impressive sweet spot. You get the broad knowledge base of an enterprise-scale 400B model without having to pay the compute toll of a fully dense network during generation.

Keep in mind that while serving 17B active parameters is fast, loading a 397B sparse model still requires massive VRAM and disk bandwidth. It is a powerful engine for developers with dedicated clusters, but remains out of reach for a modest local rig.