# Atlas with Gemma 3 12B (Ollama): 128K Context and Screenshots in 2026 > Gemma 3 12B (Ollama) pairs a 128K tokens (131,072) context with multimodal image input from an 8.1GB download, Free (self-hosted). Gemma 3 12B (Ollama) is Gemma 3 at 8.1GB with a 128K tokens (131,072) context and multimodal input, priced Free (self-hosted). Inside Atlas it fixes the one thing that made Gemma 2 unusable for agents, the 8K window, and adds image understanding, so screenshots of a failing UI can go straight into the Atlas prompt. Roughly 10GB to serve fits a 12GB card. It is general-purpose, so a 14B code specialist will produce better diffs at a similar footprint. ## Key takeaways - Gemma 3 12B (Ollama) is Free (self-hosted): 8.1GB weights, roughly 10GB to serve, fits a 12GB card. - 128K tokens (131,072) of context, sixteen times the 8K of Gemma 2, at roughly half the download size of gemma2:27b. - Multimodal from the 4b tag upward, so screenshots of a failing UI go straight into the Atlas prompt. - General-purpose, so a 14B code specialist will produce better diffs at a similar footprint. - Vision plus a long context inflates memory well past the 10GB weight floor, so budget headroom. ## Can you send a screenshot to Atlas with Gemma 3 12B? Yes. Gemma 3 12B (Ollama) is multimodal from the 4b tag upward, so it reads screenshots and diagrams rather than just text. Inside Atlas that means a screenshot of a failing UI can go straight into the prompt alongside the 128K tokens (131,072) of code context the model can already hold. Image input changes the shape of a debugging session. Instead of describing a broken layout in prose, you hand Gemma 3 12B the picture and the relevant source. Atlas indexes code by AST declarations using tree-sitter, not blind line windows, so the code half of that prompt arrives as whole components rather than arbitrary line ranges, and the model can line up what it sees on screen with what the declarations say should happen. Atlas is a terminal-native TUI rendered with SolidJS through the OpenTUI renderer, so the workflow stays in the terminal even though the input is visual. The multimodal path is one of the two reasons to pick gemma3:12b over gemma2:27b. ## How much bigger is Gemma 3 12B's context than Gemma 2's? Gemma 3 12B (Ollama) offers 128K tokens (131,072), sixteen times the 8K of Gemma 2, at roughly half the download size of gemma2:27b. That single change is what made the Gemma line usable for Atlas agent work, because an 8K window cannot survive a loop that injects tool output every turn. Atlas fans out work to subagents that can run in the foreground or in parallel background sessions, and each subagent transcript grows with every tool call. At 128K tokens (131,072), Gemma 3 12B holds a long plan-and-build session without constant compaction. The download is 8.1GB and it takes roughly 10GB to serve, which fits a 12GB card, so the sixteenfold window increase comes with a smaller file than the previous generation's flagship. Be honest about the cost though: vision plus a long context inflates memory well past the 10GB weight floor, so a full 128K session with images is not a 10GB session. ## Is Gemma 3 12B good enough to write Atlas patches? Gemma 3 12B (Ollama) is general-purpose, so a 14B code specialist will produce better diffs at a similar footprint. Inside Atlas, the documented pattern is to use gemma3:12b as the planning and review model and switch to a coder tag with /models for the build phase. Atlas drafts a plan in a read-only plan agent and asks before switching to a build agent, and that boundary is exactly where the model handoff belongs. Gemma 3 12B plans well: the 128K tokens (131,072) window lets it read broadly before it commits to an approach, and its multimodal input lets a screenshot inform that plan. The build phase is different work. Atlas computes a unified diff for every file edit and surfaces it for approval before writing, and a code-specialized model will put fewer bad diffs in front of you. Atlas lets you switch the active model and provider on the fly with favorites and recents, so the handoff is a keystroke. ## What hardware does Gemma 3 12B need for Atlas? Gemma 3 12B (Ollama) is 8.1GB of weights and roughly 10GB to serve, which fits a 12GB card. Pull the mid size with ollama pull gemma3:12b, serve Ollama locally, and check the model resolves with atlas models ollama before you rely on it in a session. The 10GB figure is a weight floor, not a ceiling. Vision plus a long context inflates memory well past it, so plan headroom if you intend to combine screenshots with a large slice of the 128K tokens (131,072) window. Add gemma3:12b to the ollama provider models map in atlas.json with limit.context 131072 and limit.output 8192 so Atlas sizes prompts correctly. Run the codebase index against the local Ollama embedder so a 128K window can be packed with your own code without any of it being uploaded, which is the point of running Gemma 3 12B locally in the first place rather than calling a hosted model with a similar window. ## When should you pick a different model than Gemma 3 12B? Pick a code specialist over Gemma 3 12B (Ollama) when writing patches is the whole job, since a 14B code specialist will produce better diffs at a similar footprint. Keep Gemma 3 12B when you want 128K tokens (131,072), image input, and a 10GB serve on a 12GB card in one free model. Gemma 3 12B is the right default for the Atlas plan-then-build flow's planning half, for UI debugging where a screenshot is the fastest description of the bug, and for anyone upgrading from gemma2:27b who wanted the window rather than the parameters. It is the wrong choice as the sole model in a heavy refactoring workflow. Because Gemma 3 12B is Free (self-hosted), running it alongside a coder tag costs disk space and nothing else, and the /models switch in the Atlas TUI makes the pairing practical rather than theoretical. ## Setup 1. Pull the mid size: ollama pull gemma3:12b (8.1GB). 2. Add gemma3:12b to the ollama provider models map in atlas.json with limit.context 131072 and limit.output 8192. 3. Serve Ollama locally and check the model resolves with atlas models ollama. 4. Use it as the planning and review model in the Atlas plan-then-build flow, switching to a coder tag with /models for the build phase. 5. Run the codebase index against the local Ollama embedder so a 128K window can be packed with your own code without any of it being uploaded. ## FAQ ### how to use gemma 3 12b with atlas Run ollama pull gemma3:12b (8.1GB), add gemma3:12b to the ollama provider models map in atlas.json with limit.context 131072 and limit.output 8192, then check it resolves with atlas models ollama. ### what is gemma 3 12b's context window Gemma 3 12B (Ollama) has a 128K tokens (131,072) context, sixteen times the 8K of Gemma 2, at roughly half the download size of gemma2:27b. ### can gemma 3 read screenshots Yes. Gemma 3 is multimodal from the 4b tag upward, so Gemma 3 12B reads screenshots and diagrams rather than just text, and a screenshot of a failing UI can go straight into the Atlas prompt. ### how much vram does gemma3:12b need Gemma 3 12B is 8.1GB of weights and roughly 10GB to serve, which fits a 12GB card. Vision plus a long context inflates memory well past that 10GB weight floor. ### is gemma 3 12b good at writing code Gemma 3 12B is general-purpose, so a 14B code specialist will produce better diffs at a similar footprint. Use it as the Atlas planning and review model and switch to a coder tag with /models for the build phase. ### is gemma 3 12b free to run Yes. Gemma 3 12B (Ollama) is Free (self-hosted). The only cost is hardware: an 8.1GB download and roughly 10GB to serve. ### gemma 3 12b vs gemma 2 27b for atlas Gemma 3 12B fixes the one thing that made Gemma 2 unusable for agents, the 8K window, and adds image understanding. It offers 128K tokens (131,072) at roughly half the download size of gemma2:27b. --- Canonical HTML: https://runatlas.sh/resources/models/gemma3-12b-local Source of truth: aeo_pages row `/resources/models/gemma3-12b-local` (segment: Models) (this file is generated from it, never hand-edited). Licence: Atlas is proprietary with a free core. It is not open source and there is no public source repository.