AI Potluck
Contribute
Model components / Fine-tuned / chat models

gpt-oss

OpenAI

OpenAI's open-weight reasoning models, released August 2025 in two sizes: gpt-oss-120b (117B parameters, 5.1B active) sized for a single 80GB GPU, and gpt-oss-20b (21B parameters, 3.6B active) for local and lower-latency use. Both are MoE models post-trained for reasoning, function calling and agentic tool use, with reasoning effort configurable at the prompt, and both require the harmony response format.

OpenAI's first open-weight release since GPT-2: Apache-2.0 weights with no training data and no training pipeline, which is the standard open-weights-closed-recipe rung. gpt-oss-safeguard is the safety tune built on this line and is scored separately in safeguards. Verified 2026-09-01 via the gpt-oss-120b HF model card, the openai/gpt-oss repository and the model-card paper (arXiv 2508.10925).

Openness

3 high confidence
3.0
weights
open(gpt-oss-120b/20b on HF, ungated safetensors)
data
closed
code
partial(inference reference implementations (PyTorch/Triton) in openai/gpt-oss
license
Apache-2.0(OSI)

Apache-2.0 weights, downloadable ungated on the Hub in both sizes, with no released training data (the card carries no datasets key and no training-data section beyond the paper's description) and only inference reference code in the openai/gpt-oss repository. Open weights with a closed recipe is the category's center of gravity, which is what 3 records. The formula lands on the fallthrough: weights open, license osi, data closed.

  • https://huggingface.co/api/models/openai/gpt-oss-120b recorded 2026-09-01

    `gated: false`, `private: false`, cardData.license `apache-2.0`, safetensors weight shards listed; 5,376,994 downloads in the trailing 30 days.

  • https://huggingface.co/openai/gpt-oss-120b/raw/main/README.md recorded 2026-09-01

    Raw card front matter records only `license: apache-2.0` with no datasets key; the body markets "Permissive Apache 2.0 license", configurable reasoning effort and fine-tunability, and carries no training-data section — the recipe is not released.

  • https://api.github.com/repos/openai/gpt-oss recorded 2026-09-01

    `license: Apache-2.0`, 20,362 stars, description "gpt-oss-120b and gpt-oss-20b are two open-weight language models by OpenAI"; the repo is reference implementations and tooling, not a training pipeline, which holds `code` at partial.

  • https://arxiv.org/abs/2508.10925 recorded 2026-09-01

    The gpt-oss-120b & gpt-oss-20b Model Card paper (OpenAI, 2025). Describes training at a high level; no corpus or pipeline release accompanies it. Corroborates data:closed.

Adoption

5 high confidence
5.0

11,896,046 downloads in the trailing 30 days across the two declared artifacts (openai/gpt-oss-20b 6,519,052; openai/gpt-oss-120b 5,376,994), which bands at level 5 (>10M) on the model adoption scale. The only open-weight line on the map at level 5 alongside qwen and gemma; the 20b SKU carries slightly more reach than the 120b.

Capability

4 medium confidence
4.0

A strong open-weight reasoner, clearly above the non-reasoning open mid-field (Llama 4 Maverick at 14.5 on the same index) and clearly below the closed frontier the category's 5s hold (gpt-5, claude-opus, gemini). That is the deepseek-r1 rung: an open-weight reasoning line competitive on math/reasoning benchmarks without leading the category, so it sits at, not above, R1's 4.

  • https://artificialanalysis.ai/models/gpt-oss-120b recorded 2026-09-01

    AA model page for gpt-oss-120b (high): embedded index data reads artificialAnalysisIntelligenceIndex 24.1 against a leaderboard whose frontier entries read ~60 (Claude Opus 5 max 63.1, GPT-5.6 Sol 60.9, Kimi K3 59.7) and whose Llama 4 Maverick entry reads 14.5; page summary: "amongst the leading models in intelligence, but somewhat expensive when comparing to other open weight models of similar size".

Verified 2026-09-01