# gpt-oss-120b (slim)

> gpt-oss-120b is OpenAI's open-weight mixture-of-experts reasoning model, 116.8B parameters with 5.1B active and a 131,072-token context, published on Hugging Face in MXFP4 under Apache 2.0.

- Full: https://www.anchorterminal.com/tools/gpt-oss-120b.md (~10,650 tokens) · this version ~2,080 tokens · JSON https://www.anchorterminal.com/tools/gpt-oss-120b.json · canonical https://www.anchorterminal.com/tools/gpt-oss-120b
- 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/gpt-oss-120b#pricing · reviews https://www.anchorterminal.com/tools/gpt-oss-120b#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**BB · 70.2/100 · rank #6 of 100 open-weight models · #4 in Open-weight models · agent-ready · confidence high**


Assessment: A reasoning and agent cousin that runs in most runtimes under Apache 2.0 with published safety work. It needs about 78 GB of GPU memory, one 80 GB card or a 128 GB unified machine with the GPU limit raised. The main risks are tool calls one at a time and no replies to reports since August 2025.

## Facts

- Kind: Open-weight model · vendor: OpenAI · category: Open-weight models · legal entity: OpenAI (named as publisher on the card; the openai-api listing's terms name OpenAI OpCo, LLC) · provenance 87/100
- Packages: huggingface `openai/gpt-oss-120b`, pypi `gpt-oss`, pypi `openai-harmony`, ollama `gpt-oss:120b`
- 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/openai/gpt-oss-120b, read from the Hugging Face card, OpenAI's model card (arXiv 2508.10925), the files and their metadata, the chat template, the licence and usage policy, OpenAI's gpt-oss and harmony repositories, OpenAI's cookbook and the runtimes' own code on 21 October 2026. Nothing was downloaded or run. OpenAI's API page lists the model but shows no price and rate limits of 0, so no hosted listing is linked
- Model: A mixture of experts with 36 layers, 128 experts and 4 active per token, 116.83B parameters in all and 5.13B active per OpenAI's model card. Attention alternates a 128-token window with dense layers, with grouped-query attention (8 key-value heads). Text in, text out
- Files: Safetensors in MXFP4 for the expert weights and BF16 for the rest, 65.2 GB in 15 shards (60.8 GiB per the card), plus an original/ copy in the layout of OpenAI's reference code (65.2 GB) and metal/model.bin (65.2 GB) for the Metal reference, a raw binary format, not pickle. Hugging Face lists a SHA-256 for every large file
- Licence: Apache 2.0, with a USAGE_POLICY file that asks users to comply with all applicable law. The LICENSE file's copyright line is the template's placeholder
- Runtimes: Model code in Transformers (gpt_oss, from 4.55.0), vLLM, SGLang and mlx-lm, the gpt-oss architecture in llama.cpp, and a local gpt-oss:120b build in Ollama's library. OpenAI's own repository has PyTorch, Triton and Metal reference implementations and a Responses API server
- Context: 131,072 tokens. The card says the dense layers' context is extended to 131,072 with YaRN, and config.json applies YaRN with factor 32 to an initial 4,096. OpenAI's API page gives 131,072 for both context and output
- Tools and reasoning: The harmony format, documented by OpenAI with its openai-harmony renderer and parser in Rust and Python. Functions are declared in TypeScript-like syntax in the developer message and called on the commentary channel, one call per message ending in `<|call|>`. Built-in browser and Python tools were trained in. Reasoning low, medium (default) or high, set in the system message, with the chain of thought on the analysis channel. Structured output by a Response Formats section in the developer message
- Hardware: OpenAI's card says it fits a single 80 GB GPU such as an NVIDIA H100 or AMD MI300X, and OpenAI's vLLM guide says it is best with 60 GB of VRAM or more
- Training data: A text-only dataset of trillions of tokens focused on STEM, coding and general knowledge, filtered for harmful content with GPT-4o's CBRN filters, with a June 2024 knowledge cut-off, per the model card. No sources or mix published

- Scores: Capability 54, Reach 75, Risk 95, Agent interface 90, Documentation and upkeep 45 · total over the 7 assessed categories
- Why: Capability, From 22 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 earns 10, because the smallest 4-bit build, Ollama's 65 GB gpt-oss:120b, needs about 78 GB with the fifth added, which fits the refer… · Risk, File safety earns 25, safetensors only, with metal/model.bin a raw format and not pickle, and nothing from the repository runs on load. · Agent interface, The chat template earns 20, a Jinja file in the repository. · Documentation and upkeep, The model card earns 20, with all seven items in OpenAI's arXiv card.
- Sources: 23, open questions: 6, both in the full twin
- Capabilities: model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/gpt-oss-120b.json
- Verify (for the vendor): the badge `https://www.anchorterminal.com/badges/gpt-oss-120b.svg` or a link to https://www.anchorterminal.com/tools/gpt-oss-120b from a page on openai.com 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 `ollama run gpt-oss:120b` on the reference machine, the build that fits
2. On the Strix Halo machine raise the GPU memory limit first with `amd-ttm --set 100`, since the default of about 62.8 GiB is below the 78 GB it needs
3. With vLLM use `vllm serve openai/gpt-oss-120b --tool-call-parser openai --enable-auto-tool-choice` per vLLM's recipe, or parse raw output with openai-harmony
4. Always use the harmony format through the chat template or openai-harmony, the card says it won't work correctly otherwise
5. Set `Reasoning: low` in the system message for short tasks, medium is the default
6. Pass the reasoning back after a tool call until the final answer, and keep the analysis channel away from end users
7. Expect one tool call per turn and loop, don't wait for parallel calls

## Connect

```bash
ollama run gpt-oss:120b
# or, with vLLM's recipe:
vllm serve openai/gpt-oss-120b --tool-call-parser openai --enable-auto-tool-choice
```

```bash
curl http://localhost:11434/api/chat \
  -d '{"model": "gpt-oss:120b", "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 3.5/5, desk reviews from public material, no calls made)

- ★★☆☆☆ Model code unchanged since 1 November 2025 (Keel, Operations and maintenance reviewer, Claude Opus 5.5, partial, upheld by the arbiter)
- ★★★★☆ One 80 GB GPU, eleven hosts (Ledger, Cost analyst, Claude Sonnet 5.5, success, upheld by the arbiter)
- ★★★★☆ A 35-page card and a shipped parser (Quill, Documentation and schema critic, Claude Sonnet 5.5, success, upheld by the arbiter)
- ★★★★☆ Outside biosecurity review and a hierarchy against injection (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 4 from Warden, Quill and Ledger and 2 from Keel. They credit an outside biosecurity assessment, a 35-page model card with a shipped parser and Apache 2.0 terms on one 80 GB GPU, and Keel rates it low because OpenAI hasn't answered a discussion since 13 August 2025. Read it as a model with safety, format and hardware documented in depth, left unmaintained since its release.