GPT-OSS 120B (local via Ollama) is the large OpenAI open-weight model. At 120B it needs serious hardware locally, but it is the strongest reasoning model you can run without sending a single token off your machine. It carries a 131,072 token context, is free to self-host, and has hosted fallbacks on identical weights at $0.15 / $0.60 per Mtok via Groq and $0.35 / $0.75 via Cerebras.
Why run GPT-OSS 120B when the code cannot leave the building?
GPT-OSS 120B is the strongest fully offline reasoning option for Atlas, which matters when the code you are working on legally cannot leave the building. In 2026, defense, health, and finance teams face exactly that constraint, and GPT-OSS 120B is the model that answers it.
An air gap changes the calculus completely. Once no token may cross the network boundary, the comparison is not GPT-OSS 120B against a frontier API, it is GPT-OSS 120B against the other models you can run on premises, and it wins that comparison on reasoning. Atlas completes the picture on the retrieval side: Atlas can build its code index with local Ollama embeddings, keeping code off third-party servers. Set the codebase index embedding provider to `ollama` so indexing stays offline too, and the entire Atlas loop, model and vectors, runs inside your perimeter.
What hardware does GPT-OSS 120B need to run locally?
GPT-OSS 120B ships as a 65GB download at OpenAI's native MXFP4 quantization and wants roughly 67GB of memory to serve, which means a multi-GPU rig or a high-memory Mac Studio. That is the price of admission for offline reasoning at 120B in 2026.
Do not attempt GPT-OSS 120B on a single consumer card. Pull it with `ollama pull gpt-oss:120b` and expect a 65GB transfer, then confirm the machine can actually serve it: `atlas device` reports whether Ollama and a GPU were detected. Register the model under the ollama provider block in atlas.json with baseURL http://localhost:11434/v1. If `atlas device` comes back without a GPU, the 67GB memory requirement will land on system RAM and the session will be unusable rather than merely slow.
How large a context can GPT-OSS 120B hold with reasoning on?
GPT-OSS 120B holds a 131,072 token context with reasoning enabled, so it can hold a large Atlas codebase-search result set and actually think about it. That combination, big window plus active reasoning, is what separates GPT-OSS 120B from local models that merely accept long prompts.
The distinction matters. Plenty of models advertise a large window and then skim it. GPT-OSS 120B is sized to reason over what it reads. Because Atlas searches code with hybrid semantic and keyword retrieval fused by reciprocal rank fusion, and indexes code by AST declarations using tree-sitter rather than blind line windows, the 131,072 tokens Atlas hands the model are dense with real declarations. A reasoning model at 120B extracts far more from that payload than a 20B one does.
When should you pick a different model than GPT-OSS 120B?
Pick a different model than GPT-OSS 120B when you want an interactive Atlas session to feel fast. Local inference throughput on GPT-OSS 120B is a fraction of Groq's, so the TUI feels sluggish, and the 65GB download plus roughly 67GB of memory rules out most workstations entirely.
If nothing forces the code to stay on the machine, the hosted route on the same weights is simply better: $0.15 / $0.60 per Mtok on Groq buys you speed you cannot reproduce on a multi-GPU rig, and $0.35 / $0.75 on Cerebras is the other option. If the hardware is not there at all, drop to a model sized for the machine you have. Atlas lets you switch the active model and provider on the fly with favorites and recents, so the local-versus-hosted decision does not have to be permanent.
Setup
- 01Run `ollama pull gpt-oss:120b` (65GB download at OpenAI's native MXFP4 quantization).
- 02Register the model under the ollama provider block in atlas.json with baseURL http://localhost:11434/v1.
- 03Confirm the machine can serve it: `atlas device` reports whether Ollama and a GPU were detected.
- 04Select GPT-OSS 120B in /models, and set the codebase index embedding provider to `ollama` so indexing stays offline too.
- 05For CI runners without a GPU, point at the identical hosted weights on Groq ($0.15 / $0.60) or Cerebras ($0.35 / $0.75).
Frequently asked questions
- what is the best offline reasoning model for coding
- GPT-OSS 120B. It is the strongest fully offline reasoning option Atlas can run, with a 131,072 token context and reasoning enabled, and it sends no tokens off your machine.
- how much memory does gpt-oss 120b need
- The MXFP4 build of GPT-OSS 120B is a 65GB download and wants roughly 67GB of memory to serve, which means a multi-GPU rig or a high-memory Mac Studio. Run `atlas device` to confirm Ollama and a GPU were detected.
- how much does gpt-oss 120b cost on groq vs cerebras
- GPT-OSS 120B is $0.15 / $0.60 per Mtok via Groq and $0.35 / $0.75 via Cerebras. Both serve identical weights to the local Ollama build, so an air-gapped workstation and a CI runner can share one model choice.
- how do i install gpt-oss 120b for atlas
- Run `ollama pull gpt-oss:120b` (65GB at OpenAI's native MXFP4 quantization), register it under the ollama provider block in atlas.json with baseURL http://localhost:11434/v1, confirm with `atlas device`, then select it in /models.
- is gpt-oss 120b too slow for interactive use
- Locally, often yes. Local inference throughput on GPT-OSS 120B is a fraction of Groq's, so an interactive Atlas session feels sluggish. If nothing requires the code to stay on the machine, the hosted Groq route at $0.15 / $0.60 per Mtok is faster.
- can atlas index my code without sending it anywhere
- Yes. Set the codebase index embedding provider to `ollama` so Atlas builds its code index with local embeddings. Paired with GPT-OSS 120B served locally, no token of your code leaves the machine.
- what context window does gpt-oss 120b have
- GPT-OSS 120B has a 131,072 token context window with reasoning enabled, which is enough to hold a large Atlas codebase-search result set and actually reason over it.
Try Atlas in your terminal
The terminal-native AI coding agent. Free core, single binary.
Install AtlasRelated guides
Atlas for Zig: A Terminal-Native AI Coding Agent for build.zig Projects in 2026
Atlas is a terminal-native AI coding agent for Zig in 2026. It reads build.zig and comptime blocks, tracks your allocators, runs zig build test behind a prompt, and runs zig fmt.
Rename a Symbol Across the Repo with Atlas in 2026
How to rename a symbol across a repo with Atlas in 2026: findReferences gets the true reference set, grep catches strings and docs, and edit refuses ambiguous matches.
Plan a Multi-File Change Before Editing with Atlas in 2026
How to plan a multi-file change with Atlas in 2026: the plan agent denies all edit tools, you research with codebase_search and lsp, then plan_exit hands off.
Atlas for Rust in 2026
Adopt Atlas, the terminal-native AI coding agent, for Rust development in 2026. Tackle borrow checker errors and clippy lints with Atlas's secure, approval-gated assistance.
Atlas vs Warp: Choosing Your AI Coding Agent in 2026
Compare Atlas, the terminal-native AI coding agent, with Warp, a smart terminal with AI Agent Mode, for developers in 2026. Evaluate features, privacy, and workflow.
Atlas for Phoenix in 2026
Atlas is a terminal-native AI coding agent for Phoenix in 2026. It reads contexts, LiveView modules, and Ecto changesets, then runs mix test behind a prompt.
Refactor a Legacy Module with Atlas in 2026
How to refactor a legacy module with Atlas in 2026: findReferences maps every callsite, apply_patch refuses to apply against a drifted file, and bash proves behavior.
Upgrade a Dependency and Fix the Breakage with Atlas (2026 Workflow)
How to upgrade a dependency and fix the breakage with Atlas in 2026: bash drives the package manager, webfetch pulls the release notes, edit fixes each compiler error.