Qwen3.5-397B-A17B: The Open Weights Behind Qwen3.5-Plus

What Qwen3.5-397B-A17B is

Qwen3.5-397B-A17B is Alibaba's flagship open-weight model of the Qwen3.5 generation: 397 billion total parameters with 17 billion active per token, released under Apache 2.0 and self-hostable. The repository carries the post-trained model in Hugging Face Transformers format, and Alibaba list it as working with Transformers, vLLM, SGLang and KTransformers.

Open weights and the hosted Qwen3.5-Plus

Alibaba state on the card that Qwen3.5-Plus, available through Alibaba Cloud Model Studio, is the hosted version corresponding to these weights — but with production features the open release does not have. Specifically a 1 million token context length by default, official built-in tools, and adaptive tool use. The open weights here run a 256K context window.

What Alibaba changed in this generation

The card describes a unified vision-language foundation: early fusion training on multimodal tokens, which Alibaba say reaches parity with Qwen3 while outperforming the separate Qwen3-VL models across reasoning, coding, agents and visual understanding. Architecturally it pairs Gated Delta Networks with sparse mixture-of-experts, which is what keeps a 397 billion parameter model to 17 billion active per token. Reinforcement learning was scaled across what Alibaba describe as million-agent environments with progressively harder task distributions. And language coverage expands to 201 languages and dialects. They also report near-100% multimodal training efficiency relative to text-only training, which is an infrastructure claim rather than a capability one.

Benchmarks

The card carries a comparison table against GPT-5.2, Claude 4.5 Opus, Gemini 3 Pro, Qwen3-Max-Thinking and Kimi K2.5, covering reasoning and agentic evaluations including HLE-Verified. Reading the rows the card labels, Qwen3.5-397B-A17B records LongBench v2 63.2, LiveCodeBench v6 83.6, HMMT February 2025 94.8, HMMT November 2025 92.7, AIME26 91.3 and BFCL-V4 72.9. GPT-5.2 records 54.5, 87.7, 99.4, 100, 96.7 and 63.1; Claude 4.5 Opus 64.4, 84.8, 92.9, 93.3, 93.3 and 77.5; and Gemini 3 Pro 68.2, 90.7, 97.3, 93.3, 90.6 and 72.5. These are Alibaba's own figures. On LMArena's text leaderboard Qwen3.5 397B A17b has 1441 Elo (rank 94); by category, 1491 Elo in Coding (rank 95), 1451 Elo in Math (rank 79), 1477 Elo in Expert (rank 81) and 1406 Elo in Creative Writing (rank 101).

Price

Roughly $0.60 per million input tokens through third-party providers, an aggregated figure rather than first-party. The hosted Qwen3.5-Plus has a 1M window; these open weights run 256K.

Architecture

The card lists 60 layers with a hidden dimension of 4,096, arranged as 15 repeats of three Gated DeltaNet layers followed by one gated-attention layer, each feeding a mixture-of-experts block. There are 512 experts, with 10 routed and 1 shared active per token and an expert intermediate dimension of 1,024. The gated-attention layers use 32 query heads and 2 key-value heads of dimension 256. The context is 262,144 tokens natively, extensible to 1,010,000.

The Qwen3.5 family

The Qwen3.5 open-weight releases covered on this site are the 397B-A17B flagship, the 122B-A10B and 35B-A3B mixture-of-experts models, and the dense 27B. Alibaba name hosted versions for two of them on Alibaba Cloud Model Studio: Qwen3.5-Plus for the 397B-A17B and Qwen3.5-Flash for the 35B-A3B. The models can also be tried in Qwen Chat at chat.qwen.ai.

Official Resources