What is the context window size for Atlas with Gemma 4 31B (Ollama)?

Atlas with Gemma 4 31B (Ollama) provides a substantial 256K tokens (262,144) context window. This large capacity is ideal for handling extensive codebases and complex development tasks within the Atlas environment.

What are the hardware requirements to run Gemma 4 31B (Ollama) with Atlas?

To run Gemma 4 31B (Ollama) effectively with Atlas, you will need a GPU with at least 24GB of VRAM. The model's weights are 20GB, and it requires roughly 22GB of VRAM to serve, which will nearly fill a 24GB card before the KV cache grows.

How much does it cost to use Gemma 4 31B (Ollama) with Atlas?

Gemma 4 31B (Ollama) is Free (self-hosted). This means there are no direct costs associated with using the model, beyond the initial hardware investment and electricity for running it locally.

Can Atlas use local embeddings with Gemma 4 31B (Ollama)?

Yes, Atlas can build its code index with local Ollama embeddings when using Gemma 4 31B (Ollama). This capability ensures that code retrieval, generation, and unified-diff review all execute on the same offline stack, enhancing privacy and control over your code.

What is the benefit of the 20GB weights in Gemma 4 31B (Ollama) compared to smaller Gemma 4 models?

The 20GB weights of Gemma 4 31B (Ollama) represent the top of the locally runnable Gemma 4 range, offering increased capacity. While the 26b tag is 18GB with the same 256K context window, the 31B variant provides a larger model for potentially more robust performance, though it requires more VRAM.

How does Atlas handle model switching with Gemma 4 31B (Ollama)?

Atlas lets you switch the active model and provider on the fly, supporting favorites and recents. When using Gemma 4 31B (Ollama), it is recommended to set a smaller 3B or 4B tag for the `small_model` slot to prevent the 20GB weights of Gemma 4 31B from being evicted during model changes.

Atlas with Gemma 4 31B (Ollama) in 2026

Atlas developers in 2026 can leverage Gemma 4 31B (Ollama) for powerful, local AI coding, offering a massive 256K token context window at no cost. This makes it a strong choice for privacy-conscious, resource-intensive code tasks, especially with its dedicated code-tuned variant.

What is Gemma 4 31B (Ollama) best at for Atlas?

Gemma 4 31B (Ollama) excels as a powerful, self-hosted option for Atlas in 2026, offering a 256K token context window and a dedicated 31b-coding-mtp-bf16 variant for specialized code tasks. Its 20GB weights represent the top of the locally runnable Gemma 4 range.

This model is ideal for Atlas users prioritizing local execution and large context windows for complex coding challenges. As the largest Gemma 4 tag at 20GB, it provides the full 256K token context window, which is the upper tier for the Gemma 4 line. A significant advantage for developers is the existence of a dedicated 31b-coding-mtp-bf16 variant, marking the first time the Gemma line has shipped an explicit coding build at flagship size. This code-tuned build ensures that Atlas can perform tasks like code search with Axis, the hybrid semantic and keyword code search, and generate code plans and edits with enhanced accuracy. Furthermore, Atlas can build its code index with local Ollama embeddings, ensuring that retrieval, generation, and unified-diff review all execute on the same offline stack, keeping sensitive code off third-party servers.

What are the cost and context tradeoffs for Gemma 4 31B (Ollama)?

While Gemma 4 31B (Ollama) is free to self-host, its 20GB weights and roughly 22GB serving requirement mean a 24GB GPU card will be nearly full, limiting KV cache growth for its 256K token context. The model offers a 262,144 token context window.

The primary tradeoff for Gemma 4 31B (Ollama) is its significant VRAM footprint. At roughly 22GB to serve, a 24GB consumer GPU card is nearly full before the KV cache even begins to grow. While the model boasts a 256K token context window, the practical capacity for the KV cache alongside the model weights on a single 24GB card will be constrained. This means that while the theoretical context window is vast, the usable portion might be limited by available VRAM for the cache during intensive operations. It is important to note that the 26b tag of Gemma 4 is 18GB with the same 256K window, indicating that the extra 2GB in the 31B variant buys capacity, not additional context. Despite these VRAM considerations, the model's pricing is Free (self-hosted), making it an attractive option for developers with suitable hardware who want to avoid recurring API costs.

When should I pick a different model for Atlas?

Consider a different model if your hardware cannot accommodate Gemma 4 31B (Ollama)'s roughly 22GB VRAM requirement, or if you need a smaller model for Atlas's `small_model` slot to avoid evicting its 20GB weights. This model is best suited for systems with at least 24GB of VRAM.

You should consider an alternative model if your local machine lacks a GPU with at least 24GB of VRAM, as Gemma 4 31B (Ollama)'s 20GB weights and 22GB serving requirement will strain lesser hardware. If you frequently switch models or need a more lightweight option for Atlas's `small_model` slot, a smaller 3B or 4B tag should be chosen to prevent the larger 20GB Gemma 4 31B weights from being evicted from VRAM. While Gemma 4 31B (Ollama) offers a powerful 256K token context, if your typical coding tasks do not require such a large context or if you prioritize faster inference over maximum context and local execution, other models might offer a better balance of performance and resource usage. Atlas lets you switch the active model and provider on the fly, so you can easily experiment with different configurations.

Setup

011: Pull the flagship Gemma 4 31B model from Ollama: `ollama pull gemma4:31b` (this is a 20GB download). For code-specific work, evaluate the `gemma4:31b-coding-mtp-bf16` variant listed on the same page.
022: Confirm roughly 22GB of free VRAM is available on your system using `atlas device` before attempting to serve the model.
033: Register `gemma4:31b` in your `atlas.json` configuration under the `ollama` provider. Set `limit.context` to `262144` and `limit.output` to `16384` to match its capabilities.
044: Configure Atlas to use `ollama/gemma4:31b` as the main model. Set the `small_model` to a smaller 3B or 4B tag to ensure the cheap slot does not evict the 20GB weights of Gemma 4 31B.
055: Run the codebase index through Ollama embeddings. This ensures that code retrieval, generation, and unified-diff review all execute on the same offline stack, leveraging Atlas's ability to index code by AST declarations using tree-sitter.

Frequently asked questions

What is the context window size for Atlas with Gemma 4 31B (Ollama)?: Atlas with Gemma 4 31B (Ollama) provides a substantial 256K tokens (262,144) context window. This large capacity is ideal for handling extensive codebases and complex development tasks within the Atlas environment.
What are the hardware requirements to run Gemma 4 31B (Ollama) with Atlas?: To run Gemma 4 31B (Ollama) effectively with Atlas, you will need a GPU with at least 24GB of VRAM. The model's weights are 20GB, and it requires roughly 22GB of VRAM to serve, which will nearly fill a 24GB card before the KV cache grows.
Is there a code-specific version of Gemma 4 31B (Ollama) for Atlas?: Yes, Google publishes a `31b-coding-mtp-bf16` variant of Gemma 4 31B. This is the first time the Gemma line has shipped an explicit coding build at flagship size, making it highly suitable for Atlas's code-centric operations.
How much does it cost to use Gemma 4 31B (Ollama) with Atlas?: Gemma 4 31B (Ollama) is Free (self-hosted). This means there are no direct costs associated with using the model, beyond the initial hardware investment and electricity for running it locally.
Can Atlas use local embeddings with Gemma 4 31B (Ollama)?: Yes, Atlas can build its code index with local Ollama embeddings when using Gemma 4 31B (Ollama). This capability ensures that code retrieval, generation, and unified-diff review all execute on the same offline stack, enhancing privacy and control over your code.
What is the benefit of the 20GB weights in Gemma 4 31B (Ollama) compared to smaller Gemma 4 models?: The 20GB weights of Gemma 4 31B (Ollama) represent the top of the locally runnable Gemma 4 range, offering increased capacity. While the 26b tag is 18GB with the same 256K context window, the 31B variant provides a larger model for potentially more robust performance, though it requires more VRAM.
How does Atlas handle model switching with Gemma 4 31B (Ollama)?: Atlas lets you switch the active model and provider on the fly, supporting favorites and recents. When using Gemma 4 31B (Ollama), it is recommended to set a smaller 3B or 4B tag for the `small_model` slot to prevent the 20GB weights of Gemma 4 31B from being evicted during model changes.

Try SeaShell in your terminal

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

Install SeaShell

Related guides

Atlas vs Qwen Code: Terminal AI Coding Agents in 2026

Atlas vs Qwen Code in 2026: Compare terminal-native AI coding agents. Atlas offers local control, permission-gated tools, and a TUI. Qwen Code features 1M token context and desktop automation.

Atlas for Julia: A Terminal-Native AI Coding Agent for Project.toml Packages in 2026

Atlas is a terminal-native AI coding agent for Julia in 2026. It reads dispatch signatures and Project.toml [deps], fixes type instabilities, runs Pkg.test(), and applies JuliaFormatter.

Atlas vs Cline: Terminal AI Coding Agents in 2026

Atlas and Cline, top AI coding agents in 2026, present distinct workflows. Atlas is terminal-native with permission-gated tools. Cline, a VS Code agent, offers Plan/Act modes and checkpoint rollbacks.

Atlas for Python in 2026

Atlas is a terminal-native AI coding agent for Python in 2026. Run it in a repo with a pyproject.toml or requirements.txt and review every diff before it lands.

Atlas for Astro: Islands, Content Collections, and Zero JS by Default in 2026

Atlas is a terminal-native AI coding agent for Astro in 2026. It reads astro.config.mjs, src/pages, and content collection schemas, drops needless client:load directives, and runs astro check.

Atlas vs Mistral Vibe for Code: Terminal AI Coding Agents in 2026

Compare Atlas and Mistral Vibe for Code in 2026. Atlas offers terminal-native TUI, permission-gated tools, and local embeddings. Mistral Vibe provides a four-model stack and EU data sovereignty.

Atlas vs JetBrains AI Assistant: Terminal AI Coding Agents in 2026

Comparing Atlas, the terminal-native AI coding agent, with JetBrains AI Assistant, the IDE-integrated AI layer, for developers in 2026. Explore pricing, local models, and workflow.

Atlas vs Amp: Choosing Your Terminal AI Coding Agent in 2026

Compare Atlas and Amp, two leading terminal AI coding agents for 2026. Atlas offers a free core and TUI, while Amp features Oracle and remote Orbs with pay-as-you-go pricing.

Browse this resource hub