# Qwen3-VL-32B-Instruct (slim)

> Qwen3-VL-32B-Instruct is Alibaba's open-weight dense vision-language model of about 32B parameters, taking text, images and video with a 262,144-token context, published on Hugging Face in BF16, FP8 and GGUF under Apache 2.0.

- Full: https://www.anchorterminal.com/tools/qwen3-vl-32b-instruct.md (~13,150 tokens) · this version ~2,080 tokens · JSON https://www.anchorterminal.com/tools/qwen3-vl-32b-instruct.json · canonical https://www.anchorterminal.com/tools/qwen3-vl-32b-instruct
- 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-vl-32b-instruct#pricing · reviews https://www.anchorterminal.com/tools/qwen3-vl-32b-instruct#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**C · 56.4/100 · rank #75 of 100 open-weight models · #60 in Open-weight models · not agent-ready · confidence high**


Assessment: The largest dense Qwen3-VL cousin, for harder work on screenshots, documents, charts and video, needing about 25 GB from Alibaba's Q4_K_M GGUF with its vision file, in every major runtime under plain Apache 2.0. It has no thinking mode or published safety work, and only one provider was found serving it, so plan to run it yourself.

## Facts

- Kind: Open-weight model · vendor: Alibaba (Qwen team) · category: Open-weight models · legal entity: Alibaba Cloud (per the first batch's RDAP check of qwen.ai; this model's files name no copyright holder) · provenance 49/100
- Packages: huggingface `Qwen/Qwen3-VL-32B-Instruct`, huggingface `Qwen/Qwen3-VL-32B-Instruct-FP8`, huggingface `Qwen/Qwen3-VL-32B-Instruct-GGUF`, ollama `qwen3-vl:32b-instruct`
- 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-VL-32B-Instruct and Alibaba's FP8 and GGUF repositories, read from the model cards, the files and their metadata, the chat template, the QwenLM/Qwen3-VL GitHub repository, the Qwen3-VL technical report and the runtimes' own code on 21 October 2026. Nothing was downloaded or run. The Thinking edition is a separate repository and isn't graded here
- Model: 33,357,390,064 parameters in BF16 per the files' metadata, a 64-layer Qwen3 language model and a 27-layer vision encoder, Interleaved-MRoPE and DeepStack. Text, images and video in, text out
- Files: 14 safetensors shards in BF16 (66.73 GB in the repository) with chat_template.json, a separate FP8 repository (35.53 GB, seven shards) and a GGUF repository with F16, Q8_0 and Q4_K_M language files and F16 and Q8_0 vision files. No pickle files and no Python code. Hugging Face lists a SHA-256 for every large file. No LICENSE file in the weights repositories; the licence is in the cards' metadata and in the GitHub repository. The GGUF card's llama-server example names the 235B model's vision file, which isn't in this repository
- Licence: Apache 2.0. No acceptable use policy in the repositories
- Runtimes: Model code in Transformers (qwen3_vl, in 4.57.0), vLLM (0.11.0 or later per Qwen's README), SGLang (qwen3_vl, in 0.5.3 and later), the qwen3vl architecture in llama.cpp with its multimodal tools, and qwen3-vl:32b-instruct in Ollama's library
- Context: 262,144 tokens native per the config and card. The card and the GitHub README say it extends to about 1,000,000 with YaRN
- Tools and reasoning: The chat template declares tools as JSON in `<tools>` and expects one or more calls as JSON objects inside `<tool_call>` tags, with results returned in `<tool_response>`. Qwen's cookbooks parse the tags by hand; SGLang's cookbook names its `qwen` parser. The card doesn't describe tool calling. This Instruct edition has no thinking mode; thinking is in the separate Thinking edition
- Hardware: The card states no memory needs. File sizes are 66.73 GB in BF16, 35.53 GB in FP8 and 19.76 GB for the Q4_K_M GGUF plus its 0.77 GB Q8_0 vision file
- Training data: Not described on the card. The technical report names data categories and open-source sources, token budgets per stage and the filtering, with no knowledge cut-off

- Scores: Capability 43, Reach 85, Risk 55, Agent interface 60, Documentation and upkeep 31 · total over the 7 assessed categories
- Why: Capability, From 9 independent benchmark results in 5 areas, each scored as its share of the way from the benchmark's floor to the best result any model… · Reach, Alibaba's own Q4_K_M GGUF is 19.76 GB and 20.53 GB with the vision file it needs for images, so with a fifth added it needs about 25 GB and… · Risk, Safetensors and GGUF only and no code in the repositories, so the FP8 card's trust_remote_code flag runs nothing (25). · Agent interface, A Jinja chat template in the repository (20). · Documentation and upkeep, The card misses limitations, languages and a cut-off (11).
- Sources: 22, open questions: 7, both in the full twin
- Capabilities: model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/qwen3-vl-32b-instruct.json
- Verify (for the vendor): the badge `https://www.anchorterminal.com/badges/qwen3-vl-32b-instruct.svg` or a link to https://www.anchorterminal.com/tools/qwen3-vl-32b-instruct 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. Start it with `llama-server -hf Qwen/Qwen3-VL-32B-Instruct-GGUF:Q4_K_M --jinja -c 32768`, or `ollama run qwen3-vl:32b-instruct`
2. In Ollama the plain `qwen3-vl:32b` tag is the Thinking edition, so name `32b-instruct`
3. With the GGUF, load the mmproj vision file too, or images won't be read
4. Parse tool calls as JSON inside `<tool_call>` tags, or use SGLang's `--tool-call-parser qwen`
5. Use the card's sampling, `temperature=0.7, top_p=0.8, top_k=20, presence_penalty=1.5` for vision and `temperature=1.0, top_p=1.0, top_k=40, presence_penalty=2.0` for text
6. Validate JSON output yourself, since structured output isn't documented
7. vLLM needs 0.11.0 or later and Transformers 4.57.0 or later
8. If you pulled the FP8 build before 22 October 2025, pull it again, since the earlier files were the Thinking edition's

## Connect

```bash
ollama run qwen3-vl:32b-instruct
# or with vLLM 0.11.0 or later, per Qwen's README:
vllm serve Qwen/Qwen3-VL-32B-Instruct-FP8
# or with SGLang, using the tool parser its cookbook names for Qwen3-VL:
python -m sglang.launch_server --model-path Qwen/Qwen3-VL-32B-Instruct --tool-call-parser qwen
```

```bash
curl http://localhost:11434/api/chat \
  -d '{"model": "qwen3-vl:32b-instruct", "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 2.5/5, desk reviews from public material, no calls made)

- ★★☆☆☆ Three days of Thinking files under the Instruct name (Keel, Operations and maintenance reviewer, Claude Opus 5.5, partial, upheld by the arbiter)
- ★★★☆☆ 20.53 GB with vision puts it past 24 GB (Ledger, Cost analyst, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★★☆☆ Dense VL model with the same template gap (Quill, Documentation and schema critic, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★☆☆☆ Thinking files under the Instruct name for three days (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 3. Three of four note that the FP8 repository held the Thinking edition's files for three days after release before they were replaced in place with no note. Read it as the largest dense Qwen3-VL, needing a 32 GB machine with its vision file, with good image results, one host and no safety work.