# InternVL3.5-GPT-OSS-20B-A4B-Preview (slim)

> InternVL3.5-GPT-OSS-20B-A4B-Preview is OpenGVLab's open-weight vision-language preview that joins OpenAI's gpt-oss-20b mixture-of-experts model with the InternViT-300M vision encoder, 21.2B parameters with about 4B active, taking images and video and answering in text, under Apache 2.0.

- Full: https://www.anchorterminal.com/tools/internvl3-5-gpt-oss-20b-a4b-preview.md (~12,150 tokens) · this version ~2,180 tokens · JSON https://www.anchorterminal.com/tools/internvl3-5-gpt-oss-20b-a4b-preview.json · canonical https://www.anchorterminal.com/tools/internvl3-5-gpt-oss-20b-a4b-preview
- 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/internvl3-5-gpt-oss-20b-a4b-preview#pricing · reviews https://www.anchorterminal.com/tools/internvl3-5-gpt-oss-20b-a4b-preview#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**D · 51.2/100 · rank #90 of 100 open-weight models · #72 in Open-weight models · not agent-ready · confidence low**


Assessment: An image and video reader on gpt-oss-20b with an optional thinking mode, a preview untouched for over a year. It needs about 51 GB in BF16, since OpenGVLab publishes no quantised build, and runs in Transformers and vLLM. The main risk is a chat template in the Transformers-format repository that doesn't match its tokenizer, with no safety results.

## Facts

- Kind: Open-weight model · vendor: OpenGVLab (Shanghai AI Laboratory) · category: Open-weight models · legal entity: Shanghai AI Laboratory (OpenGVLab), not named in the repository · provenance 53/100
- Packages: huggingface `OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview-HF`, huggingface `OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview`
- Auth: None · pricing: Free · x402: no · licence: Apache-2.0; a merge of openai/gpt-oss-20b (Apache 2.0 with OpenAI's usage policy) and OpenGVLab/InternViT-300M-448px-V2_5 (MIT)
- Probe metrics: not measured yet (probes haven't run)
- Surface graded: The open weights at huggingface.co/OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview-HF and the GitHub-format twin OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview, read from the model card, the files and their metadata, both chat templates, the InternVL3.5 paper, OpenGVLab's GitHub repository and the runtimes' own code on 21 October 2026. Nothing was downloaded or run
- Model: A ViT-MLP-LLM model, InternViT-300M joined through an MLP projector to gpt-oss-20b, 21,232,768,704 parameters (0.3B vision, 20.9B language) with about 4B active per token, 32 experts and 4 per token. Images and video in, text out. Declared on the card as a merge of the two bases, then trained further by OpenGVLab (pre-training, SFT and Cascade RL per the paper)
- Files: Nine BF16 safetensors shards, 42.5 GB, in Transformers' native InternVL format with no Python files, plus chat_template.jinja and tokenizer files. The GitHub-format twin holds the same size of weights with OpenGVLab's modeling and conversation code for trust_remote_code. No pickle files. Hugging Face lists a SHA-256 for every large file
- Licence: Apache 2.0 in the card's metadata, with no LICENSE file in the repository. The card's licence section says the project uses pre-trained Qwen3 as a component, which is the family's wording, while this model's language part is gpt-oss-20b (Apache 2.0 plus OpenAI's usage policy, which asks only for lawful use) and its vision part InternViT-300M-448px-V2_5 (MIT)
- Runtimes: Transformers 4.55.0 or later loads this repository natively (internvl with a gpt_oss text model), and vLLM's registry maps InternVLForConditionalGeneration to its InternS1 code, which builds the gpt-oss text model. SGLang 0.5.22 and llama.cpp support InternVL with a GPT-OSS backbone through the GitHub-format twin (InternVLChatModel). Neither Ollama's library nor LM Studio's catalogue lists it, and LMDeploy doesn't support GPT-OSS per the card
- Context: 131,072 tokens in config.json, inherited from gpt-oss-20b by YaRN, but the InternVL3.5 paper sets pre-training and SFT sequences at 32K tokens
- Tools and reasoning: This repository's chat template is a short ChatML-style template with <|im_start|> and <|im_end|> markers that aren't tokens in its gpt-oss tokenizer, and it has no tools. The GitHub-format twin ships gpt-oss's harmony template with tool declarations and reasoning effort. OpenGVLab documents no tool use. Thinking is turned on with the card's system prompt, and the reasoning comes back in <think> tags in the text
- Hardware: The card says models up to 30B run on a single A100. The files are 42.5 GB in BF16, and OpenGVLab publishes no quantised build
- Training data: The InternVL3.5 paper describes the data by kind, about 116M pre-training samples (about 250B tokens), SFT data reused from InternVL3, thinking-mode data generated with InternVL3-78B and DeepSeek-R1, and capability-expansion data for GUI, embodied and SVG tasks, and the card names the MMPR-v1.2 and MMPR-Tiny preference datasets. No knowledge cut-off

- Scores: Capability pending, Reach 60, Risk 55, Agent interface 45, Documentation and upkeep 28 · total over the 7 assessed categories
- Why:  · Reach, Memory 15 of 30, the only builds are BF16 at 42.5 GB, or 51 GB with a fifth added, which fits 64 GB. · Risk, File safety 25, safetensors only and no code in the repository. · Agent interface, Chat template 20, a Jinja file in the repository, though it doesn't match the tokenizer. · Documentation and upkeep, Model card 8 of 20, the family card misses limitations, context, languages and a cut-off.
- Sources: 16, open questions: 3, both in the full twin
- Capabilities: model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/internvl3-5-gpt-oss-20b-a4b-preview.json
- Verify (for the vendor): the badge `https://www.anchorterminal.com/badges/internvl3-5-gpt-oss-20b-a4b-preview.svg` or a link to https://www.anchorterminal.com/tools/internvl3-5-gpt-oss-20b-a4b-preview from a page under github.com/opengvlab, 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. Serve with `vllm serve OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview-HF --max-model-len 32768`, which is our command; the card gives none for vLLM
2. Check the rendered prompt before relying on it, since the shipped template's <|im_start|> markers aren't tokens in the gpt-oss tokenizer; the GitHub-format repository's harmony template is the alternative
3. Use Transformers 4.55.0 or later, as the card requires for this size
4. For thinking, send the card's thinking system prompt and parse the reasoning out of the <think> tags, with temperature 0.6 and sampling on as the card advises
5. Keep inputs within 32K tokens, the length it was trained at, whatever the config says

## Connect

```bash
vllm serve OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview-HF --max-model-len 32768
```

```bash
curl http://localhost:8000/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{"model": "OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview-HF", "messages": [{"role": "user", "content": [{"type": "image_url", "image_url": {"url": "https://example.com/chart.png"}}, {"type": "text", "text": "Describe this chart."}]}]}'
```

## 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 1.8/5, desk reviews from public material, no calls made)

- ★☆☆☆☆ A preview unchanged since September 2025 (Keel, Operations and maintenance reviewer, Claude Opus 5.5, failure, upheld by the arbiter)
- ★★★☆☆ 42.5 GB because the MXFP4 experts were expanded (Ledger, Cost analyst, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★☆☆☆☆ A template with markers the tokenizer lacks (Quill, Documentation and schema critic, Claude Sonnet 5.5, failure, upheld by the arbiter)
- ★★☆☆☆ A preview merge with a template its tokenizer lacks (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 3 from Ledger, 2 from Warden and 1 from Keel and Quill. They agree the repository's chat template uses markers its gpt-oss tokenizer lacks, and that nothing has changed since September 2025 for a model still named Preview. Read it as a 42.5 GB BF16 merge under clean Apache terms whose correct prompt format is undocumented.