Alibaba CloudOpen Source

Qwen3.5-397B-A17B

Compare this model

Alibaba's flagship open-weight model in the Qwen3.5 series: a 397B-parameter MoE with 17B active per token, handling text and multimodal input.

Share:XHatenaLINE

Parameters

397B (MoE, 17B active)

Context Window

TBA

License

Apache 2.0

Release Date

2026-10-26

API Pricing

API pricing for this model is not yet available

Strengths

  • •Open-weight at 397B with only 17B active per token
  • •Low adoption friction inside the trusted Qwen ecosystem
  • •Quiet launch signals fast iteration, not over-promising

Weaknesses

  • •No benchmarks published at launch
  • •17B/397B split still tracks the big model on memory and latency
  • •Multimodal quality unproven

Use Cases

  • •On-prem inference in regulated industries
  • •Cost-sensitive agents
  • •Workloads that must keep weights self-hosted

Deep Analysis

Parameters

397B (MoE, 17B active)

Context window

TBA

License

Apache 2.0 (open-weight)

Modality

Text + multimodal

Strengths

  • ・Open-weight at 397B with only 17B active per token is a genuine local-deploy story: run the smart parts without paying the full 397B bill.
  • ・It lives in the Qwen ecosystem buyers already trust for both API and self-host, so adoption friction is low.
  • ・The quiet launch suggests Alibaba is iterating fast and not over-promising on day one.

Weaknesses

  • ・Zero published benchmarks at launch. Anyone quoting a score is guessing.
  • ・A 17B/397B active split means memory and latency still track the big model unless the router is good.
  • ・Multimodal quality is unproven until vision benchmarks actually land.

Competitor Comparison

ModelGPQAPrice
Qwen3.8-27BTBAopen
DeepSeek V4.1-FlashTBAopen
Llama 4 MaverickTBAopen

Qwen3.5-397B-A17B is Alibaba's flagship open-weight model in the Qwen3.5 series: a 397B-parameter MoE with 17B active per token that handles text and multimodal input. It shipped quietly on chat.qwen.ai on Oct 26 alongside the closed Qwen3.5-Plus. No official benchmarks were published at launch, so every ranking claim is unverified until third parties run it.

Sources

Analysis generated: 2026-10-27