What is the context window size for GPT-OSS 120B (local via Ollama) in Atlas?

GPT-OSS 120B (local via Ollama) provides a 131,072 token context window, allowing Atlas to process and reason over very large codebases and search results effectively.

What are the hardware requirements to run GPT-OSS 120B (local via Ollama) locally with Atlas?

Running GPT-OSS 120B (local via Ollama) locally requires significant hardware. The MXFP4 build is a 65GB download and needs approximately 67GB of memory to serve, typically necessitating a multi-GPU rig or a high-memory Mac Studio.

Is GPT-OSS 120B (local via Ollama) free to use with Atlas?

Yes, GPT-OSS 120B (local via Ollama) is free when self-hosted. However, there are optional hosted fallbacks available via Groq at $0.15 / $0.60 per Mtok and Cerebras at $0.35 / $0.75 per Mtok.

Can Atlas use GPT-OSS 120B (local via Ollama) for offline code indexing?

Yes, Atlas can build its code index with local Ollama embeddings when using GPT-OSS 120B (local via Ollama), ensuring that all code indexing remains entirely offline and off third-party servers.

How does local inference speed compare to cloud options for GPT-OSS 120B (local via Ollama)?

Local inference throughput for GPT-OSS 120B (local via Ollama) will be a fraction of what cloud providers like Groq offer. This means interactive Atlas sessions might feel sluggish compared to using a hosted version of the model.

What are the benefits of using GPT-OSS 120B (local via Ollama) for secure development with Atlas?

GPT-OSS 120B (local via Ollama) is the strongest reasoning model you can run without sending a single token off your machine. This is crucial for secure development, especially when working on code that legally cannot leave the building, as Atlas's operations remain entirely local.

Can I use GPT-OSS 120B (local via Ollama) for both air-gapped workstations and CI runners?

Yes, GPT-OSS 120B (local via Ollama) offers hosted fallbacks on identical weights via Groq and Cerebras. This allows an air-gapped workstation to use the local model while a CI runner can use the hosted version, sharing one consistent model choice.

Atlas with GPT-OSS 120B (local via Ollama) in 2026

GPT-OSS 120B (local via Ollama) is the premier choice for Atlas users in 2026 who require the strongest reasoning capabilities without sending any code off their machine. It offers a massive 131,072 token context window, enabling deep analysis of large codebases. While self-hosted, it is free to run locally, with optional hosted fallbacks via Groq at $0.15 / $0.60 per Mtok or Cerebras at $0.35 / $0.75 per Mtok.

What is GPT-OSS 120B (local via Ollama) best at with Atlas?

For Atlas users in 2026, GPT-OSS 120B (local via Ollama) excels as the strongest fully offline reasoning model available, ensuring no code leaves your machine. Its 131,072 token context window allows Atlas to process extensive codebase-search results and perform deep analysis, crucial for sensitive projects.

GPT-OSS 120B (local via Ollama) is an OpenAI open-weight model, locally served via Ollama, making it the top choice for developers using Atlas in environments with strict data sovereignty or security requirements. Atlas leverages this model's robust reasoning to power its core capabilities, such as drafting plans in a read-only plan agent and generating unified diffs for file edits. With its substantial 131,072 token context window, Atlas can effectively utilize its hybrid semantic and keyword retrieval fused by reciprocal rank fusion to search code, then feed a large result set to GPT-OSS 120B for comprehensive understanding and action. This model also supports Atlas's ability to build its code index with local Ollama embeddings, ensuring the entire indexing process remains offline. For teams needing to maintain an air-gapped workstation while also running CI, the availability of hosted fallbacks on identical weights via Groq at $0.15 / $0.60 per Mtok or Cerebras at $0.35 / $0.75 per Mtok provides a consistent model choice across different operational needs.

What are the cost and context tradeoffs for GPT-OSS 120B (local via Ollama)?

While GPT-OSS 120B (local via Ollama) is free to run when self-hosted, its local deployment demands significant hardware. The MXFP4 build requires a 65GB download and approximately 67GB of memory to serve, necessitating a multi-GPU rig or a high-memory Mac Studio for optimal performance in 2026.

The primary tradeoff for GPT-OSS 120B (local via Ollama) is the substantial local hardware investment required. To serve the 120B parameter model, users must accommodate a 65GB download for the MXFP4 build and ensure roughly 67GB of memory is available. This typically translates to needing a multi-GPU setup or a high-memory Mac Studio, which represents a considerable upfront cost for local operation. In contrast, while the model is free when self-hosted, its local inference throughput will be a fraction of what cloud providers like Groq offer. This means an interactive Atlas session driven by a locally hosted GPT-OSS 120B might feel sluggish compared to a cloud-backed alternative. However, the benefit is a fully offline operation, where Atlas's permission-gated tool calls and indexing by AST declarations using tree-sitter can proceed without sending a single token off your machine. For scenarios where this offline capability is paramount, the hardware and performance tradeoffs are often acceptable, especially given the 131,072 token context window that allows for extensive in-memory processing.

When should you pick a different model for Atlas?

You should consider a different model for Atlas if your local hardware cannot comfortably provide the 67GB of memory required by GPT-OSS 120B (local via Ollama), or if interactive session speed is a top priority. In 2026, alternatives might offer faster inference at a lower local resource cost.

GPT-OSS 120B (local via Ollama) is not the ideal choice if your development environment lacks the robust hardware necessary to serve a 120B parameter model. Specifically, if you do not have a multi-GPU rig or a high-memory Mac Studio capable of providing approximately 67GB of memory, the local experience with Atlas will be severely degraded or impossible. Furthermore, if your workflow prioritizes rapid, interactive responses from Atlas, the local inference throughput of GPT-OSS 120B will likely feel sluggish compared to cloud-hosted models. While Atlas allows you to switch the active model and provider on the fly, if you consistently find yourself waiting for the agent to respond, a cloud-based model with faster inference, even if it means sending tokens off your machine, might be a better fit. For instance, if the legal or security constraints on code leaving your building are not absolute, leveraging the hosted fallbacks on identical weights via Groq at $0.15 / $0.60 per Mtok or Cerebras at $0.35 / $0.75 per Mtok could provide a performance boost while maintaining model consistency. However, if the core requirement is absolute offline reasoning with a large context, GPT-OSS 120B remains unparalleled.

Setup

01Pull the model: Open your terminal and execute `ollama pull gpt-oss:120b`. Be prepared for a 65GB download, as this is OpenAI's native MXFP4 quantization.
02Configure Atlas provider: Register the model under the `ollama` provider block in your `atlas.json` configuration file. Ensure the `baseURL` is set to `http://localhost:11434/v1`.
03Verify hardware capability: Run `atlas device` in your terminal. This command reports whether Ollama and a compatible GPU were detected, confirming your machine can serve the model.
04Select model in Atlas: Navigate to the `/models` interface within Atlas and select GPT-OSS 120B (local via Ollama) as your active model.
05Set embedding provider: To keep all code processing offline, set the codebase index embedding provider to `ollama` within Atlas's settings. This ensures indexing stays offline too.

Frequently asked questions

What is the context window size for GPT-OSS 120B (local via Ollama) in Atlas?: GPT-OSS 120B (local via Ollama) provides a 131,072 token context window, allowing Atlas to process and reason over very large codebases and search results effectively.
What are the hardware requirements to run GPT-OSS 120B (local via Ollama) locally with Atlas?: Running GPT-OSS 120B (local via Ollama) locally requires significant hardware. The MXFP4 build is a 65GB download and needs approximately 67GB of memory to serve, typically necessitating a multi-GPU rig or a high-memory Mac Studio.
Is GPT-OSS 120B (local via Ollama) free to use with Atlas?: Yes, GPT-OSS 120B (local via Ollama) is free when self-hosted. However, there are optional hosted fallbacks available via Groq at $0.15 / $0.60 per Mtok and Cerebras at $0.35 / $0.75 per Mtok.
Can Atlas use GPT-OSS 120B (local via Ollama) for offline code indexing?: Yes, Atlas can build its code index with local Ollama embeddings when using GPT-OSS 120B (local via Ollama), ensuring that all code indexing remains entirely offline and off third-party servers.
How does local inference speed compare to cloud options for GPT-OSS 120B (local via Ollama)?: Local inference throughput for GPT-OSS 120B (local via Ollama) will be a fraction of what cloud providers like Groq offer. This means interactive Atlas sessions might feel sluggish compared to using a hosted version of the model.
What are the benefits of using GPT-OSS 120B (local via Ollama) for secure development with Atlas?: GPT-OSS 120B (local via Ollama) is the strongest reasoning model you can run without sending a single token off your machine. This is crucial for secure development, especially when working on code that legally cannot leave the building, as Atlas's operations remain entirely local.
Can I use GPT-OSS 120B (local via Ollama) for both air-gapped workstations and CI runners?: Yes, GPT-OSS 120B (local via Ollama) offers hosted fallbacks on identical weights via Groq and Cerebras. This allows an air-gapped workstation to use the local model while a CI runner can use the hosted version, sharing one consistent model choice.

Try SeaShell in your terminal

The terminal-native AI coding agent. Free core, single binary.

Install SeaShell

Related 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: Terminal AI Coding Agents in 2026

Comparing Atlas, the terminal-native AI coding agent, with Warp, a Rust-based smart terminal with AI Agent Mode, for developers in 2026.

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.

Browse this resource hub