Qwen3.5 (397B-A17B)
used in
0
stacks
Alibaba's Apache-2.0 open-weight Qwen3.5 flagship MoE with very low active-parameter cost and broad multilingual coverage.
Qwen3.5-397B-A17B is released under Apache 2.0 with 397B total and 17B activated parameters, combining Gated Delta Networks with a sparse mixture-of-experts design (512 experts). Context is 262,144 tokens natively, extensible to about 1M tokens via YaRN, with support for 201 languages and dialects. Tool calling is supported through Qwen-Agent (including MCP configuration) and Qwen Code; it is served on Alibaba Cloud Model Studio and self-hostable via vLLM, SGLang, KTransformers, and quantized llama.cpp/Ollama builds.
Strengths
- Permissive Apache 2.0 license
- Only 17B active parameters, enabling comparatively cheap inference
- First-class MCP and tool-calling support via Qwen-Agent
- 201-language coverage for multilingual agents
Tradeoffs
- Long context beyond 262K requires YaRN extension rather than native support
- Flagship commercial Qwen3-Max capabilities are not part of the open release