Qwen3.5-397B-A17B — Open Multimodal MoE for Agents, Not Raw Scale Alone

Corrected October 14, 2026: Attributed the 8.6×–19× throughput figures to Alibaba and named the Qwen3-Max baseline and context lengths; removed an unverified blog timestamp and an unsupported Reuters attribution; corrected the pretraining description and removed an editing note left in the text.

Lunar New Year week stacked Chinese open and efficient releases. Alibaba’s proprietary close for Qwen3.5’s flagship: frontier-ish multimodal agents under Apache 2.0 with extreme sparsity—capability diffusion as the product, not only Elo screenshots.

Alibaba’s Qwen team released Qwen3.5-397B-A17B (announced February 16)—first model in the Qwen3.5 series. Open-weight, natively multimodal (vision-language via early fusion) Mixture-of-Experts: 397B total parameters, 17B active per token (high-sparsity design). Hybrid attention: Gated Delta Networks (linear attention) with sparse MoE. Pretraining: early-fusion training on trillions of multimodal tokens, multilingual coverage (201 languages/dialects; larger 250k vocabulary). Post-training: broad RL across difficult environments for agentic tool use, planning, long-horizon tasks. Vendor benches: competitive standing vs GPT-5.2, Claude 4.5 Opus, Gemini-3 Pro families across language, VL, coding agent, search agent—with Alibaba-reported decoding throughput of 8.6× (32K context) to 19× (256K context) that of Qwen3-Max. Hosted Qwen3.5-Plus: 1M context and built-in adaptive tool use. Weights: Hugging Face, ModelScope, GitHub under Apache 2.0.

Claims vs checks#

Architecture, parameter counts, Apache 2.0, and weight locations are Qwen primary. Cross-lab “competitive/leading” and throughput multiples are vendor; independent rankings had not been published at the time of writing. The 1M-token context figure applies to the hosted Plus service, which may differ from the open weights.

Limits#

• MoE serving quality depends on infrastructure; laptop demos ≠ 397B cluster reality.
• Multilingual breadth ≠ equal quality per language.
• Agentic benches are sensitive to tool scaffolds—compare like-with-like.

Sources#

• Qwen blog: “Qwen3.5: Towards Native Multimodal Agents” (2026/02/15–16). Primary.
• Reuters: “Alibaba unveils new Qwen3.5 model for ‘agentic AI era’” (Feb 16, 2026).
• Hugging Face / GitHub references in the announcement for Qwen3.5-397B-A17B.
• SemiAnalysis InferenceX: Qwen3.5-397B-A17B architecture, evaluation and inference performance (reports Alibaba’s decoding-throughput figures)
• Qwen team: QwenLM/Qwen3.5 GitHub repository (release log; archived copy March 4, 2026)
• Hugging Face: Qwen/Qwen3.5-397B-A17B model card