What it is
gpt-oss-20b is the smaller of OpenAI's two open-weight models, released in August 2025. It has 21 billion parameters with 3.6 billion active per token, and because its mixture-of-experts weights ship in 4-bit MXFP4, the whole checkpoint is about 13.8GB and runs within 16GB of memory. Its context is 131,072 tokens.
It is a text-only reasoning model with low, medium and high effort set in the system prompt, built-in tool use and a visible chain of thought, and it works correctly only with OpenAI's harmony prompt format. OpenAI says it matches o3-mini on reasoning, maths and health tests.
The weights are free under Apache 2.0, with a one-sentence usage policy asking users to follow applicable law. OpenAI does not serve it in its own API; people run it with Ollama, LM Studio, llama.cpp or vLLM, or through hosts listed on OpenRouter.
Our review
gpt-oss-20b remains a sensible default for private, text-only reasoning on modest hardware, mostly because it is fast and costs almost nothing. Artificial Analysis measures 185 output tokens a second with 0.82 seconds to first token, eighth of 64 comparable models on speed, and the cheapest route on OpenRouter charges $0.018 per million input tokens. At that rate the price of a reply barely registers.
Its intelligence is middling for the class. Artificial Analysis scores the high-effort setting at 9 on its Intelligence Index, 50th of 141 open models of similar size and just above the median of 8. It used 70 million output tokens to run that index, below the 82 million median, so unlike many reasoning models it does not spend its speed advantage on long chains of thought. Its training data stops in May 2024, which shows on anything recent.
The licence is the other reason companies keep it. OpenAI's usage policy is a single sentence asking users to obey the law, with no revenue threshold or acceptable-use list to clear with legal, and the model can be fine-tuned on consumer hardware. The Hugging Face repository logged 5.78 million downloads last month, fourteen months after release.
The constraints are fixed by design. It reads only text, the harmony format is mandatory, and the raw chain of thought it exposes is, in OpenAI's words, not intended for end users, so products need a filter between model and screen. Gemma 4's 26B A4B fits a similar memory budget at 14.4GB in Google's official 4-bit build and adds image and video input, which is where the laptop-sized slot has moved.
Where it earns its place
A developer at a law firm with a 16GB laptop
Summarise and query client documents without sending them to a cloud API
Run it in Ollama or LM Studio at medium effort; nothing leaves the machine and there is no per-token bill. Expect solid summaries but weaker reasoning than frontier APIs, and no knowledge after May 2024.
A support operations lead at a SaaS company handling 1M tickets a month
Classify and tag tickets for routing at minimal cost
On OpenRouter's cheapest route, 1M tickets at 1,000 input and 500 output tokens costs about $63 a month. Low effort keeps latency down; send the hard cases to a larger model.
An ML engineer with one workstation GPU
Fine-tune a reasoning model on internal tooling and ship it inside a product
OpenAI says the 20B can be fine-tuned on consumer hardware, and Apache 2.0 lets the result ship without a licence negotiation. The 120B sibling needs a single H100 node for the same job.
Before you commit
- Setup
- Pull it in Ollama, LM Studio or llama.cpp, or download from Hugging Face; Transformers' chat template applies harmony automatically, but direct model.generate calls must apply it themselves.
- Integrations
- Function calling, browsing and Python tool use in the harmony format; OpenAI-compatible servers through vLLM, Ollama and LM Studio; five API providers tracked by Artificial Analysis.
- Data and privacy
- Self-hosted, so prompts stay local. Apache 2.0 allows commercial use and fine-tuning; the usage policy only requires compliance with applicable law.
- Cost at scale
- Weights are free and fit a 16GB GPU. Hosted on OpenRouter from $0.018 input and $0.09 output per 1M tokens, so 1M short replies a month costs well under $100; Artificial Analysis puts it at about $0.01 per index task.
- Limits
- 131,072-token context; text only; May 2024 knowledge cutoff; three reasoning levels; output degrades without harmony formatting.
What's good
- Runs within 16GB; checkpoint about 13.8GB
- Apache 2.0 with a one-sentence usage policy
- 185 tokens a second on Artificial Analysis
- Hosted from $0.018 per 1M input tokens
What's not
- Intelligence Index of 9, mid-table for its size
- Text only; knowledge stops in May 2024
- Fails without the harmony prompt format
- Raw chain of thought needs filtering before users see it
Best for
Developers who want a free, private and fast reasoning model on a laptop or single consumer GPU, and do not need image input.
Alternatives
gpt-oss-20b: common questions
- Is gpt-oss-20b worth it?
- Fourteen months on, gpt-oss-20b is still the cheapest competent reasoning model to host, but Gemma 4 now does more in the same 16GB. We rate it 3.8 out of 5.
- How much does gpt-oss-20b cost?
- Free weights under Apache 2.0. Hosted from $0.018 input / $0.09 output per 1M tokens on OpenRouter; Artificial Analysis lists $0.07 / $0.25.
- What are the downsides of gpt-oss-20b?
- Intelligence Index of 9, mid-table for its size; text only; knowledge stops in May 2024; fails without the harmony prompt format; raw chain of thought needs filtering before users see it.
- Who is gpt-oss-20b best for?
- Developers who want a free, private and fast reasoning model on a laptop or single consumer GPU, and do not need image input.
- How much does gpt-oss-20b cost for a team or in production?
- Weights are free and fit a 16GB GPU. Hosted on OpenRouter from $0.018 input and $0.09 output per 1M tokens, so 1M short replies a month costs well under $100; Artificial Analysis puts it at about $0.01 per index task.
- Is gpt-oss-20b safe for company data?
- Self-hosted, so prompts stay local. Apache 2.0 allows commercial use and fine-tuning; the usage policy only requires compliance with applicable law.
- What are the best gpt-oss-20b alternatives?
- The closest alternatives we have reviewed are gpt-oss-120b, Gemma 4 and Phi-4-Reasoning-Vision-15B.
Answers checked Oct 22, 2026.