gpt-oss-20b
OpenAI·Aug 2025reasoningopen weights · Apache 2.0
gpt-oss-20b brings OpenAI-grade reasoning to consumer hardware: 21B total parameters, 3.6B active, runnable in 16GB of memory. It is a favorite for on-device agents, offline assistants and privacy-sensitive workloads where data cannot leave the machine. Quality sits well below the 120b sibling, but for a model that runs on a gaming laptop it is remarkably capable.
Benchmark results
No verified benchmark results tracked yet for gpt-oss-20b. This page is updated as official evaluations are published.
Where it shines
- Runs locally on 16GB consumer hardware
- Apache 2.0 — no usage restrictions
- Good tool-calling support for local agent experiments
Alternatives to gpt-oss-20b
Alibaba's trillion-parameter API flagship — frontier-adjacent quality with strong agentic tool use at mid-tier prices.
Google's November 2025 frontier breakout — 91.9% GPQA Diamond and 37.5% HLE made it the reasoning leader of its generation.
Google's 2025 workhorse flagship — first mainstream thinking model with a 1M context, still widely deployed.
The sparse-attention release that halved DeepSeek's already-lowest-in-class prices while keeping GPT-class quality.
Frequently asked questions
- How much does the gpt-oss-20b API cost?
- gpt-oss-20b costs $0.05 per million input tokens and $0.2 per million output tokens. A workload of 10M input and 1.5M output tokens per month costs about $0.8.
- What is the context window of gpt-oss-20b?
- gpt-oss-20b supports a context window of 131,072 tokens (131K), with up to 131K output tokens per response.
- Is gpt-oss-20b open source?
- Yes — gpt-oss-20b is an open-weights model released under the Apache 2.0 license, so it can be downloaded and self-hosted.
- What are the best alternatives to gpt-oss-20b?
- The closest alternatives by overall capability are Qwen3-Max, Gemini 3 Pro, Gemini 2.5 Pro, DeepSeek V3.2. See the comparison pages for detailed head-to-head breakdowns.