# Atlas with Llama 3.2 1B (local via Ollama): the 2026 offline small_model slot > Llama 3.2 1B (local via Ollama) is a 1.3GB pull with a 128,000 token window and Free (self-hosted) pricing, sized for Atlas's small_model slot rather than its build agent. Llama 3.2 1B (local via Ollama) is the cheapest way to keep Atlas's high-frequency housekeeping calls entirely offline. The pull is 1.3GB, the pricing is Free (self-hosted), and Ollama lists llama3.2:1b with a 128K window. Llama 3.2 1B is not a coding model and should not pretend to be one: at 1B parameters it cannot follow Atlas's tool schema reliably, so it belongs in the small_model slot while a hosted coding model holds the primary model slot. ## Key takeaways - `ollama pull llama3.2:1b` is 1.3GB, small enough to run on CPU alone with no GPU present. - Pricing is Free (self-hosted), which zeroes out Atlas's high-frequency small_model line on an API bill. - Ollama lists llama3.2:1b with a 128K window, so the nominal context is 128,000 tokens. - Atlas's documented config caps it at `"limit": { "context": 8192, "output": 2048 }`, well under the advertised window. - Never set Llama 3.2 1B as the primary Atlas model: 1B parameters cannot follow Atlas's tool schema reliably. - The 3B variant gives visibly better summaries for only 0.7GB more disk. ## What is Llama 3.2 1B (local via Ollama) best at inside Atlas? Llama 3.2 1B (local via Ollama) is best at Atlas's small_model traffic: titles, summaries, and other high-frequency calls that are otherwise the noisiest line on an API bill. The 1.3GB pull runs on CPU alone with no GPU present, and it costs nothing per token. Atlas separates the model that does the work from the small_model that handles cheap, constant chatter. Llama 3.2 1B (local via Ollama) is sized for the second job and only the second job. Meta ships it through Ollama as a 1.3GB download, small enough to run on CPU alone with no GPU present, which means a developer on a thin laptop can still push every summary call off a metered API and onto local silicon. Zero marginal cost for Atlas's high-frequency small_model traffic is the whole argument for the model, and it is a real one on a busy repo where those calls fire constantly. ## How big is the Llama 3.2 1B context window in Atlas? Llama 3.2 1B (local via Ollama) advertises 128,000 tokens, because Ollama lists llama3.2:1b with a 128K window. That is a large nominal window for a 1B model, so it can at least read what Atlas hands it, but Atlas's documented config caps context at 8192 for a reason. The documented Atlas setup for Llama 3.2 1B adds llama3.2:1b to the ollama models map with `"limit": { "context": 8192, "output": 2048 }`. The gap between the advertised 128,000 token window and the 8192 limit Atlas registers is deliberate: a 1B model given an enormous window still has 1B parameters of reasoning behind it, and the KV cache for a long context on a CPU-only box is slow. Set the limit Atlas documents, keep the small_model calls short, and treat the 128K figure as headroom rather than a target. ## Is Llama 3.2 1B good enough to be the primary Atlas model? No. Llama 3.2 1B (local via Ollama) must never be set as the primary Atlas model. At 1B parameters it cannot follow Atlas's tool schema reliably, and Atlas is a tool-driven agent: every tool call is permission-gated against allow, ask, and deny rules before it runs, and a model that malforms calls simply stalls. Atlas drafts a plan in a read-only plan agent and asks before switching to a build agent, then computes a unified diff for every file edit and surfaces it for approval before writing. Both of those flows assume the model can emit structured tool calls turn after turn. Llama 3.2 1B (local via Ollama) cannot hold that discipline. Quality on summaries is also visibly worse than the 3B variant for a saving of only 0.7GB, so if you have the disk and the RAM, the 3B is the better small_model. Pick the 1B when 0.7GB genuinely matters or when the machine has no GPU at all. ## How do you build a hybrid local plus hosted Atlas setup with Llama 3.2 1B? The documented hybrid is one line of atlas.json: set `"small_model": "ollama/llama3.2:1b"` and point `"model"` at a hosted coding model. Llama 3.2 1B (local via Ollama) then serves only the cheap traffic, entirely offline, while the hosted model does the actual editing. That split is the reason Llama 3.2 1B (local via Ollama) earns a slot at all. The expensive, judgment-heavy work goes to a capable hosted coder. The constant, low-stakes work goes to a 1.3GB local model that costs nothing and never leaves the machine. Atlas lets you switch the active model and provider on the fly with favorites and recents, so you can flip the primary model without touching the small_model binding. Atlas can also build its code index with local Ollama embeddings, keeping code off third-party servers, which pairs naturally with a local small_model if privacy is part of your reason for running Llama 3.2 1B. ## When should you pick a different model than Llama 3.2 1B? Pick a different model any time Llama 3.2 1B (local via Ollama) would be doing real engineering. A 1B model cannot drive Atlas's plan-then-build flow. Reach for the 3B Llama variant for better summaries, or a hosted coding model for the primary model slot, and keep the 1B strictly on cheap traffic. Llama 3.2 1B (local via Ollama) has exactly two things going for it: `ollama pull llama3.2:1b` is 1.3GB, and it is free. Run `atlas device` to see whether a GPU is present. If a GPU is present, you have better local options and should use them for the model slot. If no GPU is present, Llama 3.2 1B is still the model to use for small_model anyway, because it is the only tier that runs acceptably on CPU. The honest framing is a floor, not a recommendation: this is the cheapest thing that works for housekeeping, and nothing more. ## Setup 1. Run `ollama pull llama3.2:1b` (1.3GB) to fetch the smallest useful Llama size. 2. Run `atlas device` to see whether a GPU is present; if not, Llama 3.2 1B is the model to use anyway. 3. In atlas.json, add `llama3.2:1b` to the ollama models map with `"limit": { "context": 8192, "output": 2048 }`. 4. Set `"small_model": "ollama/llama3.2:1b"` in atlas.json so only cheap Atlas traffic runs locally. 5. Point `"model"` at a hosted coding model, giving you a hybrid setup where the primary model does the editing and the local 1B does the housekeeping. 6. Build the Atlas code index with local Ollama embeddings so retrieval vectors stay on the machine alongside the local small_model. ## FAQ ### how much does llama 3.2 1b cost to run with atlas Llama 3.2 1B (local via Ollama) is Free (self-hosted). You pay for electricity and 1.3GB of disk after `ollama pull llama3.2:1b`, and nothing per token, which is why it suits Atlas's high-frequency small_model traffic. ### can i run llama 3.2 1b without a gpu Yes. The `ollama pull llama3.2:1b` download is 1.3GB, small enough to run on CPU alone with no GPU present. Run `atlas device` to confirm whether a GPU exists on your machine; if none does, Llama 3.2 1B is the local model to use anyway. ### what context window does llama3.2:1b have in ollama Ollama lists llama3.2:1b with a 128K window, so the nominal context is 128,000 tokens. Atlas's documented config still registers it with `"limit": { "context": 8192, "output": 2048 }`, because a 1B model gains little from a very long prompt. ### should llama 3.2 1b be my main model in atlas No. Llama 3.2 1B (local via Ollama) has 1B parameters and cannot follow Atlas's tool schema reliably, so it must never be set as the primary model. Set it as `"small_model": "ollama/llama3.2:1b"` and point `"model"` at a hosted coding model. ### llama 3.2 1b vs 3b for atlas small_model The 3B variant produces visibly better summaries than Llama 3.2 1B for a saving of only 0.7GB on the 1B side. Choose the 1B when 0.7GB of disk or memory genuinely matters, or when the machine has no GPU at all. ### how do i set a local small_model in atlas.json Add `llama3.2:1b` to the ollama models map in atlas.json with `"limit": { "context": 8192, "output": 2048 }`, then set `"small_model": "ollama/llama3.2:1b"`. Leave `"model"` pointed at a hosted coding model for the actual editing work. ### is a hybrid local and cloud atlas setup possible Yes. Atlas lets you switch the active model and provider on the fly with favorites and recents, and the documented hybrid pins `"small_model"` to ollama/llama3.2:1b while `"model"` stays on a hosted coding model, so only cheap traffic stays local. --- Canonical HTML: https://runatlas.sh/resources/models/llama-3-2-1b-local Source of truth: aeo_pages row `/resources/models/llama-3-2-1b-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.