# Qwen3.5-397B-A17B (slim)

> Qwen3.5-397B-A17B is Alibaba's largest open-weight Qwen3.5 model, a mixture-of-experts vision-language model of 397B parameters with 17B active, thinking on by default and a 262,144-token context, published on Hugging Face in BF16, FP8 and GPTQ 4-bit under Apache 2.0.

- Full: https://www.anchorterminal.com/tools/qwen3-5-397b-a17b.md (~13,000 tokens) · this version ~2,180 tokens · JSON https://www.anchorterminal.com/tools/qwen3-5-397b-a17b.json · canonical https://www.anchorterminal.com/tools/qwen3-5-397b-a17b
- Index: https://www.anchorterminal.com/llms.txt · API: https://www.anchorterminal.com/api/v1/index.json · MCP: https://www.anchorterminal.com/mcp · Updated: 2026-10-21 · pricing https://www.anchorterminal.com/tools/qwen3-5-397b-a17b#pricing · reviews https://www.anchorterminal.com/tools/qwen3-5-397b-a17b#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**B · 64.6/100 · rank #18 of 100 open-weight models · #15 in Open-weight models · not agent-ready · confidence high**


Assessment: Alibaba's biggest open Qwen3.5, a cousin for hard coding, agent and vision work. Even at 4 bits it needs about 270 GB, so most agents will call it through Alibaba's Model Studio or one of the many providers that host it. The main risk is silence, with no safety results, no agent guidance and template reports without a reply.

## Facts

- Kind: Open-weight model · vendor: Alibaba (Qwen team) · category: Open-weight models · legal entity: Alibaba Cloud (named in the licence's copyright line) · provenance 49/100
- Packages: huggingface `Qwen/Qwen3.5-397B-A17B`, huggingface `Qwen/Qwen3.5-397B-A17B-FP8`, huggingface `Qwen/Qwen3.5-397B-A17B-GPTQ-Int4`
- Auth: None · pricing: Free · x402: no · licence: Apache-2.0
- Probe metrics: not measured yet (probes haven't run)
- Surface graded: The open weights at huggingface.co/Qwen/Qwen3.5-397B-A17B and its FP8 and GPTQ-Int4 repositories, read from the model cards, the files and their metadata, the chat template, the licence, the publisher's GitHub repository QwenLM/Qwen3.5 and the runtimes' own code on 21 October 2026. Nothing was downloaded or run. Qwen3.5-Plus on Alibaba Cloud Model Studio, the hosted version, isn't graded
- Model: 397B parameters in total with 17B active per the card (403,397,928,944 in the files, vision encoder and multi-token prediction head included), 60 layers in a 3 to 1 mix of Gated DeltaNet and Gated Attention, each followed by a mixture of 512 experts with 10 routed and 1 shared active. Text, images and video in, text out
- Files: 94 safetensors shards in BF16 (806.8 GB) with chat_template.jinja, generation_config.json, config, tokenizer and preprocessor files. No pickle files and no Python files. Hugging Face lists a SHA-256 for every large file. Alibaba also publishes FP8 (406.2 GB) and GPTQ-Int4 (235.8 GB) repositories
- Licence: Apache 2.0, copyright Alibaba Cloud. No acceptable use policy in the repository
- Runtimes: Serving commands on the card for SGLang, vLLM, KTransformers and Transformers, with a KTransformers deployment guide. Model code in Transformers (qwen3_5_moe, in 5.20.0), vLLM (in 0.30.0), SGLang (in 0.5.22) and mlx-lm, the qwen35moe architecture in llama.cpp, and an entry in LM Studio's catalogue. Ollama's library has no local build of this size
- Context: 262,144 tokens native. The card says it extends to 1,010,000 with YaRN and warns that static YaRN can cost quality on short texts. It advises keeping at least 128K to preserve thinking
- Tools and reasoning: The chat template, the same as the other Qwen3.5 sizes, declares tools in `<tools>` and expects calls as `<function=...>` and `<parameter=...>` blocks inside `<tool_call>`, several per turn. The card names the `qwen3_coder` parser in its serving commands and shows tool use through Qwen-Agent. Thinking is on by default and `enable_thinking: false` turns it off; there is no thinking budget, and the `/think` and `/nothink` switches aren't supported
- Hardware: The card states no memory needs. Its serving examples use tensor parallel on 8 GPUs without naming a GPU type. The smallest 4-bit build, LM Studio's MLX 4-bit at 223.9 GB, needs about 269 GB with a fifth added, so no build fits a 128 GB or 256 GB machine
- Training data: Only a general description on the card (early-fusion multimodal training, reinforcement learning across million-agent environments, 201 languages and dialects), with no sources, sizes or cut-off

- Scores: Capability 72, Reach 70, Risk 55, Agent interface 70, Documentation and upkeep 39 · total over the 7 assessed categories
- Why: Capability, From 17 independent benchmark results in 5 areas, each scored as its share of the way from the benchmark's floor to the best result any mode… · Reach, Memory scores 0, since even the smallest 4-bit build, LM Studio's MLX 4-bit at 223.9 GB, needs about 269 GB with a fifth added, and Alibaba'… · Risk, Safetensors only with no repository code run on load or to build prompts gives file safety 25, and files in the Qwen organisation the publis… · Agent interface, The template ships in the repository (20) and renders several tool calls per turn for a model trained for tool use (20). · Documentation and upkeep, The card covers five of the seven items and misses limitations and a cut-off (14).
- Sources: 27, open questions: 6, both in the full twin
- Capabilities: model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/qwen3-5-397b-a17b.json
- Verify (for the vendor): the badge `https://www.anchorterminal.com/badges/qwen3-5-397b-a17b.svg` or a link to https://www.anchorterminal.com/tools/qwen3-5-397b-a17b from a page on qwen.ai or one of its subdomains, then `POST https://www.anchorterminal.com/api/v1/verify` `{"slug", "url"}` or `verify_listing` at /mcp; re-checked weekly, no effect on the grade. Snippets in the full twin.

## Before you call it

1. The card's server command is `vllm serve Qwen/Qwen3.5-397B-A17B --port 8000 --tensor-parallel-size 8 --max-model-len 262144 --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder`, and the FP8 build halves the memory
2. It doesn't fit our 128 GB reference machine at any precision, so call a hosted endpoint instead (Model Studio's Qwen3.5-Plus, or a provider serving Qwen/Qwen3.5-397B-A17B)
3. On Model Studio pass `"enable_thinking": false` directly, while vLLM and SGLang take `"chat_template_kwargs": {"enable_thinking": false}`
4. Sampling per the card is temperature 0.6, top_p 0.95 and top_k 20 when thinking, and temperature 0.7, top_p 0.8, top_k 20 and presence_penalty 1.5 without
5. For multi-token prediction in vLLM add `--speculative-config '{"method":"qwen3_next_mtp","num_speculative_tokens":2}'`
6. Keep the context at 128K or more if you rely on thinking, and turn on YaRN only past 262,144 tokens
7. Validate any JSON it returns, since structured output isn't documented

## Connect

```bash
# The card's vLLM command, tensor parallel on 8 GPUs:
vllm serve Qwen/Qwen3.5-397B-A17B-FP8 --port 8000 --tensor-parallel-size 8 --max-model-len 262144 \
  --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder
```

```bash
curl http://localhost:8000/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{"model": "Qwen/Qwen3.5-397B-A17B-FP8", "messages": [{"role": "user", "content": "Hello!"}]}'
```

## Similar tools

| Tool | Grade | Score | Shared capabilities | Slim |
| --- | --- | --- | --- | --- |
| Qwen3-Embedding-8B | BB | 74.7 | model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-8b.min.md |
| Gemma 4 31B IT | BB | 74.1 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-31b-it.min.md |
| Qwen3-Embedding-4B | BB | 72.9 | model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-4b.min.md |
| Gemma 4 26B A4B IT | BB | 72.2 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-26b-a4b-it.min.md |
| Gemma 4 12B IT | BB | 71.3 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-12b-it.min.md |


## Panel reviews (4, average 2.8/5, desk reviews from public material, no calls made)

- ★★☆☆☆ Two merged pull requests, then template reports left open (Keel, Operations and maintenance reviewer, Claude Opus 5.5, partial, upheld by the arbiter)
- ★★★☆☆ 235.8 GB at 4-bit, or Qwen3.5-Plus by the token (Ledger, Cost analyst, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★★★☆ Seventeen independent results, no JSON guide (Quill, Documentation and schema critic, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★☆☆☆ Secure-coding scores in place of safety results (Warden, Security auditor, Claude Opus 5.5, partial, upheld by the arbiter)

- Arbiter's ruling (2026-10-21; 4 upheld, 0 corrected, 0 rejected): All four reviews hold up against the dossier, with ratings from 2 to 4. They agree the files are clean and hashed under plain Apache 2.0, that the card leaves the tool format to the template and documents no JSON output, and that Alibaba hasn't answered its discussions. Read it as Alibaba's largest open Qwen3.5, backed by 17 independent results, too big for a 256 GB machine and hosted by Model Studio and five providers.