# Atlas with GPT-OSS 20B (local via Ollama): Local Reasoning on a 16GB Card in 2026 > GPT-OSS 20B runs on a single 16GB GPU with a 131,072 token context, and the identical weights are served by Groq at $0.075 / $0.30 per Mtok. GPT-OSS 20B (local via Ollama) is OpenAI's open-weight 20B model, designed for reasoning and agentic tasks rather than chat. It runs on a single 16GB GPU and has a 131,072 token context, which is generous for a model this size. Inside Atlas it is free self-hosted, and the identical weights are served by Groq at $0.075 / $0.30 per Mtok if you want to burst to fast hosted inference without changing models. ## Key takeaways - GPT-OSS 20B has genuine reasoning support in a 20B open-weight model, which is rare: most local models this size are completion-only. - It runs on a single 16GB GPU, the most common consumer configuration. - The context window is 131,072 tokens, generous for a model this size. - Free self-hosted, or $0.075 / $0.30 per Mtok on Groq using identical weights. - At 20B it drops context on long agent chains and is weaker at pure code generation than Qwen3-Coder or Devstral. ## What makes GPT-OSS 20B different from other 20B local models? GPT-OSS 20B has genuine reasoning support in a 20B open-weight model, which is rare: most local models this size are completion-only. OpenAI designed GPT-OSS 20B for reasoning and agentic tasks rather than chat, and that design goal is the reason to run it inside Atlas. Reasoning at 20B is unusual, and it changes what the model can do in an agent harness. GPT-OSS 20B can hold a problem, consider an approach, and then act, rather than pattern-matching the next plausible token. Atlas drafts a plan in a read-only plan agent and asks before switching to a build agent, and a reasoning-capable model makes that plan phase worth reading. Compare that with a completion-only local model, which produces a plan that is really just a restatement of the request. The distinction is why GPT-OSS 20B is a credible pick despite its size. ## Can GPT-OSS 20B run on a single 16GB GPU? Yes. GPT-OSS 20B runs on a single 16GB GPU, the most common consumer configuration in 2026. Pull it with `ollama pull gpt-oss:20b` and register it under the ollama provider in atlas.json with "limit": { "context": 131072 }. The hardware requirement is the practical argument for GPT-OSS 20B. A 16GB card is the configuration most developers actually have, and GPT-OSS 20B fits it while still offering reasoning. Setting the context limit correctly matters: register 131072 in atlas.json so Atlas does not truncate prompts before the model would. Once that is in place, select GPT-OSS 20B from /models in the TUI. Everything else in Atlas behaves as normal, including the fact that every Atlas tool call is permission-gated against allow, ask, and deny rules before it runs. ## How much does GPT-OSS 20B cost on Groq? GPT-OSS 20B is free self-hosted, and identical weights are served by Groq at $0.075 / $0.30 per Mtok. That $0.075 input rate makes GPT-OSS 20B one of the cheapest ways to burst to fast hosted inference without changing models mid-project. The local-and-hosted symmetry is the sharpest feature of GPT-OSS 20B. When your laptop is thermally throttling or you are on a CI runner with no GPU, set GROQ_API_KEY and pick openai/gpt-oss-20b. The weights are the same, so prompts that worked locally still work, and the price is $0.075 / $0.30 per Mtok. Atlas lets you switch the active model and provider on the fly with favorites and recents, so this is not a config migration, it is a switch inside a running session. ## What are the real limits of GPT-OSS 20B inside Atlas? GPT-OSS 20B is small enough at 20B parameters that it drops context on long agent chains, so it churns through Atlas permission prompts on complex tasks. GPT-OSS 20B is also weaker at pure code generation than Qwen3-Coder or Devstral, which are coding-specialized models. Honesty about GPT-OSS 20B saves a frustrating afternoon. The 131,072 token window is generous on paper, but a 20B model does not use a large window as reliably as a bigger one does. On a twenty-turn agent chain, GPT-OSS 20B loses earlier context and starts proposing tool calls that Atlas's permission gate rejects, which is exactly the churn you feel as a wall of approval prompts. Keep GPT-OSS 20B for shorter, reasoning-heavy tasks and do not ask it to carry a long refactor. ## When should you pick a different model than GPT-OSS 20B? Pick a different model than GPT-OSS 20B when the task is pure code generation. GPT-OSS 20B is weaker there than Qwen3-Coder or Devstral, which are coding-specialized, and it churns through Atlas permission prompts on complex tasks because 20B is small enough to drop context. A good split is to use GPT-OSS 20B for the reasoning it is designed for and hand the writing to a coding model. Diagnose with GPT-OSS 20B, then switch to a coding-specialized local model for the implementation pass. If the task is both long and hard, the honest answer is that 20B is not the right size, and either the hosted Groq route at $0.075 / $0.30 or a larger model is the better call. Atlas snapshots file changes as git patches so edits can be diffed and rolled back, which makes experimenting with the split cheap. ## Setup 1. Run `ollama pull gpt-oss:20b`. 2. Register it under the ollama provider in atlas.json with "limit": { "context": 131072 }. 3. Select GPT-OSS 20B from /models in the Atlas TUI. 4. To burst to hosted inference on the same weights, set GROQ_API_KEY and pick openai/gpt-oss-20b. 5. Keep a coding-specialized model available for the implementation pass, since GPT-OSS 20B is weaker at pure code generation. ## FAQ ### can you run a reasoning model on a 16gb gpu Yes. GPT-OSS 20B runs on a single 16GB GPU and has genuine reasoning support, which is rare at that size since most local models this size are completion-only. Pull it with `ollama pull gpt-oss:20b`. ### how much does gpt-oss 20b cost on groq GPT-OSS 20B is $0.075 / $0.30 per Mtok on Groq. Groq serves identical weights to the local Ollama build, so you can burst to fast hosted inference without changing models. ### what is the context window of gpt-oss 20b GPT-OSS 20B has a 131,072 token context window, which is generous for a 20B model. Register it in atlas.json with "limit": { "context": 131072 }. ### how do i set up gpt-oss 20b in atlas Run `ollama pull gpt-oss:20b`, register it under the ollama provider in atlas.json with "limit": { "context": 131072 }, then select it from /models in the TUI. For hosted inference, set GROQ_API_KEY and pick openai/gpt-oss-20b. ### is gpt-oss 20b good at writing code GPT-OSS 20B is weaker at pure code generation than Qwen3-Coder or Devstral, which are coding-specialized. It was designed for reasoning and agentic tasks rather than chat, so use it to diagnose and hand implementation to a coding model. ### why does gpt-oss 20b trigger so many permission prompts in atlas At 20B, GPT-OSS 20B is small enough that it drops context on long agent chains, so it proposes tool calls that Atlas's permission gate rejects. Keep it for shorter, reasoning-heavy tasks rather than long refactors. ### gpt-oss 20b vs gpt-oss 120b GPT-OSS 20B runs on a single 16GB GPU and is free self-hosted or $0.075 / $0.30 per Mtok on Groq. Both share the 131,072 token context, but the 20B is the one that fits mainstream consumer hardware. --- Canonical HTML: https://runatlas.sh/resources/models/gpt-oss-20b-local Source of truth: aeo_pages row `/resources/models/gpt-oss-20b-local` (segment: Models) (this file is generated from it, never hand-edited). Licence: Atlas is proprietary with a free core. It is not open source and there is no public source repository.