Qwen2.5-Coder 32B (Ollama) is the largest Qwen2.5-Coder size Ollama publishes, and the strongest fully offline coding model most people can actually run. The pull is 20GB, serving it takes roughly 22GB, which means a 24GB card (3090, 4090, 7900 XTX) or a 32GB-plus Mac. Context is 32K tokens (32,768) and pricing is Free (self-hosted). In Atlas it handles plan-then-build flows with far fewer derailments than the 7B or 14B.
What makes Qwen2.5-Coder 32B (Ollama) the strongest offline Atlas model?
Qwen2.5-Coder 32B (Ollama) is the largest Qwen2.5-Coder size Ollama publishes, and its 20GB Q4_K_M weights are the practical ceiling for a single 24GB consumer GPU. Qwen2.5-Coder 32B is the model that fills that 24GB slot best for code, which is why offline Atlas users land on it.
A 24GB card (3090, 4090, 7900 XTX) or a 32GB-plus Mac defines the outer edge of what most developers own. Qwen2.5-Coder 32B (Ollama) is the coding model that uses that budget most effectively. Inside Atlas it handles plan-then-build flows and permission-gated tool chains with far fewer derailments than the 7B or 14B. Since every Atlas tool call is permission-gated against allow, ask, and deny rules before it runs, fewer derailments translates directly into fewer approval prompts you have to reject, which is what makes a 32B feel qualitatively different from a 7B rather than merely better.
How much VRAM does Qwen2.5-Coder 32B need to serve in Atlas?
Qwen2.5-Coder 32B (Ollama) needs roughly 22GB to serve against 20GB of weights on disk. On a 24GB card that leaves about 2GB of slack, so a long context will push you into offload and tank throughput. Check headroom with `atlas device` before you commit.
The 22GB serving figure is the number that decides whether Qwen2.5-Coder 32B (Ollama) is viable for you, not the 20GB download. On a 24GB GPU the margin is thin. Raising num_ctx grows the KV cache, and once the KV cache no longer fits, the runtime spills layers to system RAM and tokens per second collapses. Run `atlas device` first: you want roughly 22GB free, and more if you raise the context. Set small_model to ollama/qwen2.5-coder:1.5b so the cheap slot does not evict 20GB of weights every time Atlas writes a session title.
Is 32K context a problem for Qwen2.5-Coder 32B in Atlas?
Qwen2.5-Coder 32B (Ollama) carries the same 32K tokens (32,768) as the 1.5B, which is small next to what modern cloud coders offer. Register it in atlas.json with limit.context 32768 and limit.output 8192, and accept that Atlas compaction will do real work on large sessions.
The honest tradeoff with Qwen2.5-Coder 32B (Ollama) is that scaling the parameters did not scale the window. You get far better reasoning inside the same 32K tokens (32,768). For Atlas that is workable because Atlas indexes code by AST declarations using tree-sitter, not blind line windows, so what enters the window is dense. Atlas also searches code with hybrid semantic and keyword retrieval fused by reciprocal rank fusion, which means the right chunks tend to arrive rather than a lot of adjacent ones. Still, if your work genuinely needs a huge window, no Qwen2.5-Coder size will give it to you.
How do you run Atlas fully air-gapped on Qwen2.5-Coder 32B?
Qwen2.5-Coder 32B (Ollama) is air-gapped by construction: weights, prompts, diffs, and, with the Ollama embedder, the vector index all stay on the machine. Set model to ollama/qwen2.5-coder:32b and point the codebase index at the local Ollama embedder to close the last hole.
Air-gapping an agent means closing four channels, not one. The weights are local once you run `ollama pull qwen2.5-coder:32b` (20GB). The prompts are local because inference is local. The diffs are local because Atlas computes a unified diff for every file edit and surfaces it for approval before writing, on your machine. The last channel is retrieval, and Atlas can build its code index with local Ollama embeddings, keeping code off third-party servers. Point the codebase index at the local Ollama embedder and the AST chunks and their vectors stay entirely offline. Qwen2.5-Coder 32B (Ollama) is Free (self-hosted), so nothing about that arrangement carries a per-token cost.
When should you not choose Qwen2.5-Coder 32B?
Do not choose Qwen2.5-Coder 32B (Ollama) without roughly 22GB of free memory. Below a 24GB card or a 32GB-plus Mac, the 14B at 9.0GB is the realistic local ceiling, and forcing the 32B into insufficient VRAM produces offload and throughput that makes Atlas unusable.
The second reason to skip Qwen2.5-Coder 32B (Ollama) is window. The 32K tokens (32,768) it offers is the same as the 1.5B, and if the blocker on your work is context rather than reasoning, spending 20GB of VRAM will not fix it. The third reason is speed: dense 32B inference on a consumer card is not fast, and if your workflow is many short turns rather than a few deep ones, a smaller Qwen2.5-Coder size may serve you better. Atlas lets you switch the active model and provider on the fly with favorites and recents, so registering more than one tag costs nothing.
Setup
- 01Pull the top size: ollama pull qwen2.5-coder:32b (20GB).
- 02Check headroom with atlas device before you commit; you want roughly 22GB free, and more if you raise the context.
- 03Add qwen2.5-coder:32b to the models map of the ollama provider in atlas.json with limit.context 32768 and limit.output 8192.
- 04Set model to ollama/qwen2.5-coder:32b so the 32B holds the build agent slot.
- 05Set small_model to ollama/qwen2.5-coder:1.5b so the cheap slot does not evict 20GB of weights.
- 06Point the codebase index at the local Ollama embedder to keep AST chunks and their vectors entirely offline.
Frequently asked questions
- can i run qwen2.5-coder 32b on a 24gb gpu
- Yes, but with thin margins. Qwen2.5-Coder 32B (Ollama) is 20GB of weights and roughly 22GB to serve, so a 24GB card (3090, 4090, 7900 XTX) works. Raise the context and you will push into offload and tank throughput.
- what is the best local model for agentic coding in 2026
- Qwen2.5-Coder 32B (Ollama) is the largest Qwen2.5-Coder size Ollama publishes and the strongest fully offline coding model most people can actually run. In Atlas it handles plan-then-build flows and permission-gated tool chains with far fewer derailments than the 7B or 14B.
- how do i check if my machine can run a 32b model with atlas
- Run atlas device before you commit. For Qwen2.5-Coder 32B (Ollama) you want roughly 22GB free, and more if you raise the context above the default. The pull itself is 20GB: ollama pull qwen2.5-coder:32b.
- does qwen2.5-coder 32b have a bigger context than the 7b
- No. Qwen2.5-Coder 32B (Ollama) has 32K tokens (32,768), the same window as the 1.5B and the 7B. Scaling to 32B parameters bought reasoning inside the same window, not a wider one.
- how do i keep atlas completely air-gapped
- Set model to ollama/qwen2.5-coder:32b, keep small_model on ollama/qwen2.5-coder:1.5b, and point the codebase index at the local Ollama embedder. Weights, prompts, diffs, and vectors then all stay on the machine.
- what small_model should i pair with a local 32b in atlas
- Set small_model to ollama/qwen2.5-coder:1.5b. Pairing a 986MB-class tag with the 32B means Atlas titles and summaries do not evict 20GB of Qwen2.5-Coder 32B weights from a 24GB card.
- is qwen2.5-coder 32b free
- Qwen2.5-Coder 32B (Ollama) is Free (self-hosted). The cost is hardware: 20GB of disk, roughly 22GB of VRAM or unified memory, and a 24GB card or 32GB-plus Mac to serve it at usable speed.
Try Atlas in your terminal
The terminal-native AI coding agent. Free core, single binary.
Install AtlasRelated guides
Atlas vs Amp: Terminal AI Coding Agents in 2026
Compare Atlas, a terminal-native AI coding agent with free core and local embeddings, against Amp, Sourcegraph's agent featuring Oracle and Orbs, for developers in 2026.
Locate Where a Behavior Is Implemented with Atlas in 2026
How to locate where a behavior is implemented with Atlas in 2026: codebase_search for meaning, grep for exact text, and the lsp tool for the symbol graph.
Atlas for Clojure: A Terminal-Native AI Coding Agent for deps.edn and Kaocha in 2026
Atlas is a terminal-native AI coding agent for Clojure in 2026. It reads deps.edn aliases and namespace requires, runs clojure -M:test with Kaocha, and applies cljfmt.
Atlas vs Claude Code: Terminal AI Coding Agents in 2026
Compare Atlas and Claude Code in 2026 for terminal AI coding. Atlas offers free core, BYO model keys, and robust diffing, while Claude Code provides deep Claude tuning and strong multi-step execution.
Atlas for Haskell in 2026
Atlas is a terminal-native AI coding agent for Haskell in 2026. Run it in a project with a .cabal file or a stack.yaml, let cabal build drive it, and review the diff.
Atlas vs Fine.dev: Terminal AI Coding Agents in 2026
Compare Atlas and Fine.dev for terminal AI coding in 2026. Atlas offers a free core TUI with permission-gated tools, while Fine.dev provides asynchronous cloud agents and prebuilt libraries.
Atlas for Fortran: fpm.toml, Explicit Interfaces, and fprettify in 2026
Atlas is a terminal-native AI coding agent for Fortran in 2026. It reads modules, explicit interfaces, and intent declarations, runs fpm test behind a prompt, and runs fprettify.
Atlas for PHP in 2026
Atlas, the terminal-native AI coding agent, empowers PHP developers in 2026 with intelligent code understanding, secure workflows, and direct integration for Composer and PSR standards.