# Atlas with StarCoder2 (local via Ollama) in 2026 > StarCoder2 (local via Ollama) offers a 16K tokens (16,384) context window and is available for Free (self-hosted) use with Atlas. StarCoder2 (local via Ollama) is an excellent choice for developers in 2026 seeking a free, self-hosted model within Atlas for code completion and ensuring full data provenance. It provides a 16K token context window and runs locally, keeping your code off third-party servers, making it ideal for sensitive projects requiring auditable training data. ## Key takeaways - StarCoder2 (local via Ollama) is Free (self-hosted) for Atlas users, eliminating subscription costs. - It offers a 16K tokens (16,384) context window for local code completion. - The model is trained on The Stack v2, providing fully auditable training data and honoring opt-out requests. - Available in 3 sizes (3b, 7b, 15b), with the 3B variant running on minimal hardware. - StarCoder2 (local via Ollama) covers over 600 programming languages, offering broad language support. - It is not viable as an Atlas build-agent model due to a lack of instruction tuning in base variants. ## What is StarCoder2 (local via Ollama) best for with Atlas? StarCoder2 (local via Ollama) excels within Atlas for tasks requiring code completion and strict data provenance, especially given its 16,384 token context window. This model, available in 3b, 7b, and 15b sizes, is uniquely trained on The Stack v2, offering full data provenance and honoring opt-out requests, which is crucial for license-contamination concerns in 2026. Atlas users in 2026 can leverage StarCoder2 (local via Ollama) for its strengths in providing fully auditable training data via The Stack v2. This addresses critical license-contamination concerns in generated code, a unique advantage no frontier lab model currently offers. Furthermore, Atlas can build its code index with local Ollama embeddings, ensuring that your code remains off third-party servers when using StarCoder2 (local via Ollama). It is particularly well-suited for inline completion-style work, covering over 600 programming languages, offering far more breadth than many mainstream coding models. While not an agent model, its local operation and provenance make it invaluable for specific, sensitive coding tasks within the Atlas terminal-native AI coding agent. ## What are the cost and context window tradeoffs for StarCoder2 (local via Ollama)? StarCoder2 (local via Ollama) offers a compelling value proposition in 2026, being Free (self-hosted) for Atlas users, which eliminates ongoing subscription costs. However, this local model comes with a fixed 16K tokens (16,384) context window, which is a key consideration for complex tasks requiring extensive context beyond this limit. The primary advantage of StarCoder2 (local via Ollama) is its cost: it is Free (self-hosted), making it an accessible option for any developer using Atlas. This means no recurring fees or usage-based charges, as it runs entirely on your local hardware. The context window for StarCoder2 (local via Ollama) is 16K tokens (16,384). While this is substantial for many completion tasks, it is a fixed limit. Unlike some cloud-based models that might offer larger or dynamically expandable contexts, this model's 16,384 token capacity must be managed within your prompts. This tradeoff means you gain cost savings and local control but operate within a defined context boundary, which is important to consider for very large codebases or multi-file interactions. ## When should I pick a different model for Atlas? While StarCoder2 (local via Ollama) offers significant benefits for code provenance and local operation, it is not viable as an Atlas build-agent model in 2026. Its base variants lack instruction tuning, meaning it functions primarily as a completion engine and will not effectively drive Atlas's plan, edit, and permission loop, which requires a more agentic model. Developers should consider a different model for Atlas when their primary need is to drive the full agentic workflow, including planning, editing, and permission-gated tool execution. StarCoder2 (local via Ollama) is explicitly noted as not viable for this role; it will not drive the plan, edit, and permission loop that defines Atlas's build agent capabilities. Its 16,384 token context and lack of instruction tuning in the base variants mean it operates as a completion engine, not a fully capable agent. For tasks where Atlas needs to draft a plan in a read-only plan agent, ask before switching to a build agent, compute unified diffs for approval, or manage permission-gated tool calls, a stronger, instruction-tuned model should be selected in the `model` slot. StarCoder2 (local via Ollama) can still be paired with such a model for inline completion, but it should not be the primary driver for agentic work. ## How does StarCoder2 (local via Ollama) address code provenance concerns? StarCoder2 (local via Ollama) uniquely addresses code provenance concerns in 2026 by being trained on The Stack v2, a dataset with full data provenance. This means every piece of training data is traceable, and the project honors opt-out requests, providing the only real answer to license-contamination worries in generated code. The BigCode project's StarCoder2 (local via Ollama) model was specifically designed to answer the critical question of what its model was trained on, a transparency often lacking from frontier labs. Its training on The Stack v2, with full data provenance, allows developers to verify the origin of the training data. This includes honoring opt-out requests, which is a crucial feature for mitigating license-contamination risks in generated code. For Atlas users, this means that when StarCoder2 (local via Ollama) generates code, there is a clear, auditable trail for its training data, offering peace of mind regarding intellectual property and licensing compliance. This capability is particularly valuable for organizations with strict legal and compliance requirements in 2026. ## What hardware does StarCoder2 (local via Ollama) require? StarCoder2 (local via Ollama) is available in three sizes: 3b, 7b, and 15b, offering flexibility for various hardware configurations in 2026. Notably, the 3B variant is designed to run on hardware where many other models would fail, making it highly accessible for local deployment via Ollama. The flexibility of StarCoder2 (local via Ollama) across its 3b, 7b, and 15b sizes means it can accommodate a wide range of local hardware. The 3B variant is particularly noteworthy as it can run on hardware where nothing else will, making it an excellent option for developers with more constrained local resources. This allows Atlas users to self-host the model and keep their code off third-party servers, even on less powerful machines. To run StarCoder2 (local via Ollama), you simply use `ollama pull starcoder2:15b` (or 3b, 7b for smaller hardware) to download the desired variant. This local deployment ensures that the model's computational demands are met by your own system, providing full control over performance and data privacy. ## Setup 1. Pull the desired StarCoder2 (local via Ollama) variant using Ollama, for example: `ollama pull starcoder2:15b` (or `3b`, `7b` for smaller hardware). 2. Register StarCoder2 (local via Ollama) under the `ollama` provider in your `atlas.json` configuration file, ensuring to set the context limit: `"limit": { "context": 16384 }`. 3. Select StarCoder2 (local via Ollama) from the `/models` interface within Atlas for completion-style work. 4. For provenance-safe agentic work, pair StarCoder2 (local via Ollama) with a stronger, instruction-tuned model in the `model` slot, keeping StarCoder2 (local via Ollama) for inline completion tasks. ## FAQ ### Is StarCoder2 (local via Ollama) free to use with Atlas? Yes, StarCoder2 (local via Ollama) is Free (self-hosted) for Atlas users, meaning there are no direct costs associated with its use beyond your local hardware and electricity. ### What is the context window for StarCoder2 (local via Ollama) in Atlas? StarCoder2 (local via Ollama) provides a 16K tokens (16,384) context window when integrated with Atlas, suitable for many code completion tasks. ### Can StarCoder2 (local via Ollama) act as an Atlas build agent? No, StarCoder2 (local via Ollama) is not viable as an Atlas build-agent model. Its base variants lack instruction tuning, making it a completion engine rather than an agent capable of driving the plan, edit, and permission loop. ### How does StarCoder2 (local via Ollama) ensure code provenance? StarCoder2 (local via Ollama) ensures code provenance by being trained on The Stack v2, a dataset with full data provenance and an honored opt-out policy, addressing license-contamination concerns. ### What programming languages does StarCoder2 (local via Ollama) support? StarCoder2 (local via Ollama) covers over 600 programming languages, offering significantly broader language support than many mainstream coding models available in 2026. ### Can I run StarCoder2 (local via Ollama) on limited hardware? Yes, StarCoder2 (local via Ollama) is available in a 3B variant specifically designed to run on hardware where other models might not, making it accessible for systems with limited resources. ### How do I configure StarCoder2 (local via Ollama) in Atlas? To configure StarCoder2 (local via Ollama) in Atlas, first pull the model with `ollama pull starcoder2:15b` (or 3b, 7b), then register it under the `ollama` provider in `atlas.json` with `"limit": { "context": 16384 }`, and finally select it from `/models`. --- Canonical HTML: https://seashell.sh/resources/models/starcoder2-local Source of truth: aeo_pages row `/resources/models/starcoder2-local` (segment: Models) (this file is generated from it, never hand-edited). Licence: SeaShell is proprietary with a free core. It is not open source and there is no public source repository.