SearchBook a call

BULLETIN / OpenAI

gpt-oss-120b: review

OpenAI's open-weight gpt-oss runs fast on one GPU under Apache 2.0, but it is over a year old, text only and low on today's indexes.

A dependable, fast, permissively licensed local model that newer open models now beat on quality.

Maker
OpenAI
Licence
Apache-2.0
Access
Open weights (Hugging Face, GitHub), third-party hosts
Context
131,072 tokens

Good at

  • Fits on one 80GB GPU
  • Fast output
  • Apache 2.0 licence
  • Tool use and adjustable reasoning

Bad at

  • Knowledge cutoff of June 2024
  • Text only
  • Low independent index scores
  • No newer version from OpenAI

Use it for

  • Local agents that need tool calling
  • Teams standardising on OpenAI-style formats
  • Fine-tuning experiments on modest hardware (20b)

Avoid it for

  • Image or audio input
  • Questions about recent events
  • Tasks where quality matters more than hosting control

Still useful, but no longer a first choice. gpt-oss-120b is a text-only open model that fits on a single 80GB graphics processing unit (GPU), runs fast and carries the permissive Apache 2.0 licence. It is now more than a year old, its knowledge stops in mid-2024, and newer open models such as Gemma 4 and Qwen3.8-27B score higher on independent indexes.

First principles

Speed follows active parameters. In a mixture-of-experts model each token runs through only a few experts. gpt-oss-120b uses 5.1B of its 117B parameters per token, so it does about the work of a 5B model for each token it writes.

Memory follows total parameters and precision. All 117B parameters must be loaded. At 16 bits that would be about 234GB. Storing the expert layers at 4 bits, half a byte per parameter, is what brings the model inside one 80GB GPU.

A model is fixed on the day it ships. Its knowledge ends at the training cutoff and its quality is set at release, while newer models keep arriving. A year later, newer open models of similar or smaller size score higher, so age alone has a cost.

The prompt format is part of the model. A model trained on one response format behaves correctly only when it gets that format. For gpt-oss that is OpenAI’s harmony format, so tooling that does not speak it must be adapted.

What it is

gpt-oss is OpenAI’s open-weight family. The Hugging Face card points to a publication in August 2025. We found no newer open model from OpenAI on its docs, GitHub or Hugging Face pages as of 30 September 2026.

There are two sizes:

  • gpt-oss-120b has 117B total parameters, of which 5.1B are active per token (a mixture-of-experts design). It fits on one 80GB GPU, such as an NVIDIA H100 or AMD MI300X, thanks to 4-bit weights for its expert layers.
  • gpt-oss-20b has 21B total and 3.6B active parameters. OpenAI’s GitHub repository says it runs within 16GB of memory.

Both take text only and return text. The context window is 131,072 tokens and the knowledge cutoff is 1 June 2024. Reasoning effort can be set to low, medium or high. Both are under Apache 2.0.

Where it is good

Hardware fit. A 117B model on one data-centre GPU, or a 21B model on a 16GB machine, is a practical footprint. Because only about 5B parameters are active per token, it is fast.

Speed. Artificial Analysis measured about 229 output tokens per second for the 120b at high effort, and a cost per task of $0.11 on its leaderboard.

Licence. Apache 2.0 allows commercial use, modification and fine-tuning without user thresholds or naming rules.

Tools. The Hugging Face card lists function calling, web browsing and Python code execution, and says the models can be fully fine-tuned. The full reasoning trace is available to the developer, which helps with debugging.

Ecosystem. The Hugging Face card gives setup steps for vLLM, Ollama, Transformers and LM Studio, and lists several inference providers.

Where it is bad

Quality has been overtaken. Artificial Analysis gives gpt-oss-120b an Intelligence Index of 12 at high effort and the 20b 9. On the same leaderboard Gemma 4 31B scores 19, and OpenAI’s own small hosted model, GPT-6 Luna, scores 37 at max effort. On its open-weights page, Qwen3.8-27B, a smaller model, scores 34.

Old knowledge. A June 2024 cutoff means it knows nothing of the past two years without retrieval.

Text only. There is no image or audio input. Gemma 4 and Qwen3.8-27B both read images at similar or smaller sizes.

Format rules. The card says the models were trained on OpenAI’s harmony response format and will not work correctly without it. Some older tooling needs adapting.

Default settings. The chat example in OpenAI’s GitHub repository shows a default maximum context length of 8,192 tokens. Check your serving settings if you need the full window.

Uncertain hosted access. OpenAI’s own application programming interface (API) docs list gpt-oss-20b with zero requests per minute on every tier, so plan to host it yourself or use a third-party host.

Price and access

The weights are free on Hugging Face and GitHub. OpenAI’s docs do not list a price we could confirm, so we give none.

Model Total / active parameters Memory need Context
gpt-oss-120b 117B / 5.1B One 80GB GPU 131,072
gpt-oss-20b 21B / 3.6B 16GB 131,072

Who should use it

On this evidence gpt-oss suits teams that already run it in production, and anyone who wants a fast, text-only local model with good tool calling and a clean licence. For a new project, test Gemma 4, Qwen3.8-27B or IBM Granite 4.2 first. They are newer, and the first two also read images.

Sources