# Atlas with IBM Granite 3.3 8B (Ollama) in 2026 > IBM Granite 3.3 8B (Ollama) provides a generous 128K token context window, making it a powerful, free, self-hosted option for Atlas users. IBM Granite 3.3 8B (Ollama) is an excellent choice for driving Atlas's small_model functions in 2026, providing reliable instruction following and long tool transcript handling at no cost. This self-hosted model offers a substantial 128K token context window, making it ideal for local summaries, titles, and subagent interactions, while keeping your code entirely off third-party servers. It is a general-purpose model, not a code specialist, but its robust instruction following makes it a solid assistant. ## Key takeaways - IBM Granite 3.3 8B (Ollama) is Free (self-hosted) and provides a 128K token context window. - It is ideal for Atlas's `small_model` slot, handling summaries, titles, and subagent chatter. - The 4.9GB model requires roughly 7GB of VRAM, making it suitable for 8GB graphics cards. - It offers reliable structured output and instruction following, crucial for Atlas's agent workflows. - As a general model, it will be outperformed by code specialists for actual diff generation. - Atlas can build its code index with local Ollama embeddings, keeping code off third-party servers. ## What is IBM Granite 3.3 8B (Ollama) best for in Atlas? IBM Granite 3.3 8B (Ollama) excels as Atlas's small_model in 2026, handling tasks like generating titles and summaries with its reliable instruction following. This 4.9GB model, requiring roughly 7GB to serve, leverages a 128K token context window to manage long tool transcripts effectively. For developers using Atlas, IBM Granite 3.3 8B (Ollama) is particularly strong in roles where robust instruction following and structured output are paramount, rather than raw code generation. It is well-suited for Atlas's small_model slot, where it can power functions such as drafting a plan in a read-only plan agent, generating concise summaries, and creating descriptive titles. Its ability to hold long tool transcripts within its 128K token context window ensures that subagent chatter and complex multi-step interactions remain coherent. Furthermore, by using local Ollama embeddings, Atlas can build its code index without sending sensitive code to third-party servers, enhancing privacy. This model also benefits from IBM enterprise licensing, simplifying procurement for organizations already using IBM products. ## Cost and Context Window Tradeoffs for IBM Granite 3.3 8B (Ollama) IBM Granite 3.3 8B (Ollama) offers a compelling value proposition in 2026, being entirely Free (self-hosted) while providing a substantial 128K tokens (131,072) context window. This 4.9GB download, requiring about 7GB of VRAM, allows long-transcript work to run efficiently on an 8GB graphics card. The primary advantage of IBM Granite 3.3 8B (Ollama) is its cost-effectiveness: it is Free (self-hosted), eliminating ongoing API expenses. Its generous 128K token context window is a significant strength, enabling Atlas to maintain extensive conversational history and tool call transcripts, which is crucial for complex coding tasks and subagent coordination. While the model itself is a 4.9GB download, it typically requires around 7GB of VRAM to serve, making it accessible for developers with an 8GB graphics card. However, a key tradeoff is that IBM Granite 3.3 8B (Ollama) is a general-purpose model, not a code specialist. While it follows instructions reliably, for tasks requiring precise diff generation or highly optimized code suggestions, an 8B code specialist model will generally outperform it. Developers must weigh the benefit of free, local operation and a large context against the need for specialized code generation capabilities. ## When to choose a different model for Atlas While IBM Granite 3.3 8B (Ollama) is a strong contender for Atlas's small_model in 2026, developers should consider alternative models for specific tasks, particularly when raw code generation is the priority. This 8B parameter model, despite its 128K context, is a generalist. Developers should consider a different model when their primary need within Atlas is the generation of highly optimized code or precise unified diffs for file edits. IBM Granite 3.3 8B (Ollama) is a general model, and while it is a well-behaved instruction follower, an 8B code specialist model will typically deliver superior results for actual diff generation. Atlas computes a unified diff for every file edit and surfaces it for approval before writing, and for these critical code modification steps, a specialized model might offer more accurate and efficient suggestions. Additionally, while the 8B version is robust, a smaller 2B tag exists at 1.5GB if extreme resource constraints are a concern, but its capability drops off sharply, making it less suitable for most Atlas workflows. For scenarios demanding peak code performance, pairing IBM Granite 3.3 8B (Ollama) as the small_model with a dedicated coder model for the main agent can provide a balanced approach. ## Setup 1. Pull the model: `ollama pull granite3.3:8b` (this is a 4.9GB download, aliased to `:latest`). 2. Declare the Ollama provider in your `atlas.json` configuration under `@ai-sdk/openai-compatible` with `baseURL http://localhost:11434/v1`. 3. Register `granite3.3:8b` with `limit.context 131072` and `limit.output 8192` to correctly configure its context window and output limits. 4. Assign `granite3.3:8b` to the `small_model` slot in Atlas so that tasks like titles, summaries, and subagent chatter run locally and free. 5. Switch the codebase index to use the Ollama embedder, ensuring the `small_model` and the vector store share the same offline runtime for enhanced privacy. ## FAQ ### Can IBM Granite 3.3 8B (Ollama) generate code for Atlas? IBM Granite 3.3 8B (Ollama) is a general-purpose model, not a code specialist. While it follows instructions well, for actual diff generation and complex code modifications within Atlas, an 8B code specialist model will typically provide superior results. It is best suited for tasks like summaries and planning. ### What is the context window for IBM Granite 3.3 8B (Ollama) in Atlas? IBM Granite 3.3 8B (Ollama) offers a substantial 128K tokens (131,072) context window. This allows Atlas to maintain long tool transcripts and extensive conversational history for complex tasks and subagent interactions. ### Is IBM Granite 3.3 8B (Ollama) free to use with Atlas? Yes, IBM Granite 3.3 8B (Ollama) is Free (self-hosted). You download the 4.9GB model and run it locally via Ollama, incurring no API costs. ### What hardware is needed to run IBM Granite 3.3 8B (Ollama) with Atlas? The IBM Granite 3.3 8B (Ollama) model is a 4.9GB download and typically requires approximately 7GB of VRAM to serve. This means it can run effectively on systems equipped with an 8GB graphics card. ### How does IBM Granite 3.3 8B (Ollama) enhance privacy for Atlas users? By running IBM Granite 3.3 8B (Ollama) locally and switching Atlas's codebase index to the Ollama embedder, all code processing and embeddings remain on your machine. This keeps your sensitive code off third-party servers, enhancing privacy. ### Can I use a smaller version of IBM Granite 3.3 with Atlas? A 2B tag of IBM Granite 3.3 exists at 1.5GB, offering a smaller footprint. However, its capability drops off sharply compared to the 8B version, making it generally less effective for the robust instruction following and long context needed for Atlas's agent workflows. ### How does IBM Granite 3.3 8B (Ollama) handle tool calls in Atlas? IBM Granite 3.3 8B (Ollama) is a well-behaved instruction follower that holds long tool transcripts. This makes it reliable for Atlas's permission-gated tool calls and for exposing Model Context Protocol servers' tools to the agent, ensuring coherent multi-step operations. --- Canonical HTML: https://seashell.sh/resources/models/granite3-3-8b-local Source of truth: aeo_pages row `/resources/models/granite3-3-8b-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.