# Atlas with Mistral NeMo 12B (Ollama): 128K Context on a 12GB Card in 2026 > Mistral NeMo 12B (Ollama) gives Atlas a 128K practical context from a 7.1GB download, roughly 9GB to serve, at Free (self-hosted) pricing. Mistral NeMo 12B (Ollama) is a joint Mistral and NVIDIA 12B at 7.1GB, built for a long window and strong multilingual coverage. Inside Atlas the pricing is Free (self-hosted), and the context window is 128K tokens (practical; Ollama's tag page renders 1000K from the model config). Treat 128K as the real figure. Roughly 9GB to serve means long-context work runs on a 10GB or 12GB card, which is the whole reason to choose this model: a genuinely long window on hardware most developers already own. ## Key takeaways - Mistral NeMo 12B (Ollama) is a 7.1GB download with a 128K practical window, roughly 9GB to serve, so long-context work runs on a 10GB or 12GB card. - The 1000K figure on Ollama's tag page comes from max_position_embeddings, not from trained long-context performance. Do not design around it. - Set limit.context 131072 in atlas.json, the 128K Mistral actually markets. - Co-developed with NVIDIA and quantization-aware trained for FP8, so it degrades gracefully under aggressive quantization. - Apache 2.0 licensed with strong multilingual and tokenizer coverage (Tekken), and Free (self-hosted). ## What is Mistral NeMo 12B (Ollama) best at inside Atlas? Mistral NeMo 12B (Ollama) is best at long-context work inside Atlas on modest hardware. A 7.1GB download with a 128K practical window, roughly 9GB to serve, means a long-context agent runs on a 10GB or 12GB card instead of a workstation. Long context is normally something you rent. The local models that hold 128K tend to want a 24GB card or better, so developers with a 12GB GPU quietly accept a 32K ceiling and work around it. Mistral NeMo 12B breaks that pattern: 7.1GB of weights, roughly 9GB to serve, and a 128K practical window. Inside Atlas that changes what you can attempt in one session, because Atlas injects retrieved code chunks, unified diffs, and tool results on every turn, and a 128K budget absorbs a multi-file refactor that a 32K model cannot hold. Mistral NeMo 12B is also Apache 2.0 licensed with strong multilingual and tokenizer coverage (Tekken), which helps on non-English identifiers and comments. ## Why does Ollama show a 1000K context for Mistral NeMo 12B? Ollama's tag page renders 1000K for Mistral NeMo 12B because that figure comes from max_position_embeddings in the model config, not from trained long-context performance. Mistral markets it as a 128K-context model, and 128K is the number to design an Atlas configuration around. The gap between a config maximum and a trained window is where local-model configurations quietly break. In atlas.json, register mistral-nemo:12b with limit.context 131072, the 128K Mistral actually markets, rather than trusting the 1000K the tag page renders. Setting 1,000,000 does not give you a million usable tokens, it gives you a model that degrades in ways that are hard to attribute, because Atlas will keep packing context that the model was never trained to attend over. Verify the resolved window with atlas models ollama after you configure it. The 1000K figure is not a lie, it is a different measurement, and treating it as a capability is the mistake. ## How much does it cost to run Atlas on Mistral NeMo 12B? Mistral NeMo 12B (Ollama) is Free (self-hosted). Budget roughly 9GB of VRAM or unified memory at moderate context, on top of a 7.1GB pull. There is no API key, no per-token bill, and no metered charge for a long 128K prompt. The cost argument is sharper for a long-context model than for a small one, because hosted providers price by the token and long context is exactly how an agent loop runs up a bill. With Mistral NeMo 12B the pricing is Free (self-hosted), so feeding Atlas a large local slice of the repository costs nothing but memory. Pair that with local Ollama embeddings on the codebase index, so a long-context model can be fed large local slices of the repo with nothing leaving the machine. Mistral NeMo 12B was co-developed with NVIDIA and quantization-aware trained for FP8, so it degrades gracefully under aggressive quantization, which matters when you are squeezing a 128K KV cache onto a 12GB card. ## When should you pick a different model than Mistral NeMo 12B? Pick a different model than Mistral NeMo 12B (Ollama) when the job is pure diff generation. Mistral NeMo 12B is a middle capacity general model, and for writing patches a 14B code specialist is usually stronger. Choose NeMo for context length, not for peak code quality. Mistral NeMo 12B is a context play, and the honest cost of that play is capability per parameter. If your Atlas sessions are mostly targeted single-file edits where the relevant code fits easily in 32K, you are paying for a window you never fill and giving up patch quality to a code specialist at a similar size. The reverse case is where NeMo wins: sprawling changes, long transcripts, large retrieved slices, and codebases where non-English identifiers and comments benefit from the Tekken tokenizer's coverage. Atlas lets you switch the active model and provider on the fly with favorites and recents, so a reasonable setup registers mistral-nemo:12b for long-horizon work and a coder tag for the build phase. ## Can Atlas feed a 128K local model without sending code to a vendor? Yes. Run the Atlas codebase index on Ollama embeddings, and a 128K model like Mistral NeMo 12B can be fed large local slices of the repo with nothing leaving the machine. Atlas can build its code index with local Ollama embeddings, keeping code off third-party servers. A long context window is only useful if you can fill it with the right code, and filling it means retrieval. Atlas searches code with hybrid semantic and keyword retrieval fused by reciprocal rank fusion, and it indexes code by AST declarations using tree-sitter, not blind line windows, so the 128K of context that Mistral NeMo 12B receives is made of whole declarations rather than truncated fragments. Running that index on the local Ollama embedder means the retrieval half of the loop is as private as the inference half. Atlas still computes a unified diff for every file edit and surfaces it for approval before writing, and it snapshots file changes as git patches so edits can be diffed and rolled back. ## Setup 1. Pull it: ollama pull mistral-nemo:12b (7.1GB, aliased to :latest) 2. In atlas.json register mistral-nemo:12b with limit.context 131072 (the 128K Mistral actually markets) rather than trusting the 1000K the tag page renders 3. Budget roughly 9GB of VRAM or unified memory at moderate context 4. Set model to ollama/mistral-nemo:12b and verify the resolved window with atlas models ollama 5. Run the codebase index on Ollama embeddings so a long-context model can be fed large local slices of the repo with nothing leaving the machine ## FAQ ### what is the real context window of mistral nemo 12b 128K tokens is the practical figure Mistral markets. Ollama's tag page renders 1000K, but that comes from max_position_embeddings in the model config, not from trained long-context performance. ### how much VRAM does mistral nemo 12b need Roughly 9GB of VRAM or unified memory at moderate context, from a 7.1GB download, so a 10GB or 12GB card can serve it. ### what limit.context should I set for mistral-nemo:12b Set limit.context 131072 in atlas.json, the 128K Mistral actually markets, rather than trusting the 1000K the Ollama tag page renders. Verify with atlas models ollama. ### is mistral nemo good for non-english code comments Yes. Mistral NeMo 12B is Apache 2.0 licensed with strong multilingual and tokenizer coverage (Tekken), which helps on non-English identifiers and comments. ### does mistral nemo quantize well Yes. Mistral NeMo 12B was co-developed with NVIDIA and quantization-aware trained for FP8, so it degrades gracefully under aggressive quantization. ### best local model with long context for a coding agent Mistral NeMo 12B is a strong option at 7.1GB with a 128K practical window and Free (self-hosted) pricing, though for pure diff generation a 14B code specialist is usually stronger. ### how do I run mistral nemo with atlas Run ollama pull mistral-nemo:12b, register it in atlas.json with limit.context 131072, set model to ollama/mistral-nemo:12b, and verify the resolved window with atlas models ollama. --- Canonical HTML: https://runatlas.sh/resources/models/mistral-nemo-12b-local Source of truth: aeo_pages row `/resources/models/mistral-nemo-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.