# Atlas with Qwen3 14B (Ollama): the local planning model for 2026 > Qwen3 14B (Ollama) is a dense 9.3GB model with 40K tokens (40,960) of context, Free (self-hosted), and it suits Atlas planning sessions more than raw diff generation. Qwen3 14B (Ollama) is dense Qwen3 at 9.3GB, the general-purpose counterpart to Qwen2.5-Coder 14B. Context is 40K tokens (40,960) and pricing is Free (self-hosted). Qwen3 14B is better at prose, planning, and mixed reasoning, and slightly less specialized on raw code completion, which makes it a good pick when Atlas is doing architecture work rather than pure edits. ## Key takeaways - 9.3GB Q4_K_M download, roughly 11GB to serve, the same hardware slot as Qwen2.5-Coder 14B. - 40K tokens (40,960) of context, wider than the 32K Qwen2.5-Coder line, useful for Atlas plan agent transcripts. - Dense architecture with no MoE routing makes throughput predictable on a single consumer GPU. - Free (self-hosted), so keeping both a planner and a coder tag registered costs disk, not money. - A general model, not a code specialist: Qwen2.5-Coder 14B is usually better on pure diff generation. ## When should you use Qwen3 14B (Ollama) instead of a coder model? Use Qwen3 14B (Ollama) when Atlas is doing architecture work rather than pure edits. Qwen3 14B is better at prose, planning, and mixed reasoning than Qwen2.5-Coder 14B at the same 9.3GB footprint, and its 40K tokens (40,960) window holds a longer plan transcript. Atlas drafts a plan in a read-only plan agent and asks before switching to a build agent, and those two phases reward different models. Planning is reasoning and prose: what should change, in what order, and why. Qwen3 14B (Ollama) is a general model, so it does that well. Generating the actual unified diff is a code specialist's job, and Qwen2.5-Coder 14B is usually better on pure diff generation at the same size. The documented Atlas setup embraces the split: assign Qwen3 14B as the model for planning sessions and keep a coder tag registered alongside it. ## How wide is the Qwen3 14B context window in Atlas? Qwen3 14B (Ollama) has 40K tokens (40,960), wider than the 32K of the Qwen2.5-Coder line, which is useful for Atlas plan agent transcripts. Register it in the ollama provider models map in atlas.json with limit.context 40960 and limit.output 8192. The extra 8,192 tokens over the Qwen2.5-Coder window sounds small and is not. An Atlas plan agent transcript accumulates: the retrieved AST chunks, the reasoning, the proposed steps, the tool results. Qwen3 14B (Ollama) at 40K tokens (40,960) absorbs more of that before compaction kicks in. The honest counterweight is that 40K is modest next to the 256K the qwen3 MoE tags advertise, so if window is the binding constraint on your work, a dense 14B is not the answer no matter how good its reasoning is. ## Why does dense architecture matter for Qwen3 14B throughput? Qwen3 14B (Ollama) is dense, with no MoE routing, which makes throughput predictable on a single consumer GPU. Every parameter of the 9.3GB Q4_K_M weights fires on every token, so tokens per second stays steady across prompts instead of varying with which experts a router picks. Predictable throughput is underrated when Atlas is running long sessions. A mixture-of-experts model can be fast on average and uneven in practice. Qwen3 14B (Ollama) is dense, so a given prompt length produces roughly the same latency every time, which makes it easier to reason about how long an Atlas plan-then-build run will take. The 9.3GB Q4_K_M download wants roughly 11GB to serve, so it shares hardware requirements with Qwen2.5-Coder 14B and you can swap between the two on the same 12GB or 16GB card without rethinking your memory budget. ## How do you switch between Qwen3 14B and a coder model mid-session in Atlas? Register both tags in atlas.json, then use `/models` in the Atlas TUI to switch mid-session. Atlas lets you switch the active model and provider on the fly with favorites and recents, so Qwen3 14B (Ollama) can plan and a Qwen2.5-Coder tag can execute inside the same session. The documented Atlas setup for Qwen3 14B (Ollama) is explicit about this: assign it as the model for planning sessions and keep a coder tag registered alongside it, then use `/models` in the Atlas TUI to switch mid-session. Because Qwen3 14B and Qwen2.5-Coder 14B share the roughly 11GB serving requirement, swapping between them on a single GPU is a tag change rather than a hardware change. Atlas is a terminal-native TUI rendered with SolidJS through the OpenTUI renderer, so the model switcher is a keystroke, not a config edit and a restart. ## What are the real tradeoffs of running Qwen3 14B locally? Qwen3 14B (Ollama) is a general model, not a code specialist, and Qwen2.5-Coder 14B is usually better on pure diff generation at the same 9.3GB size. The 40K tokens (40,960) window is also modest next to the 256K on the qwen3 MoE tags. Both tradeoffs on Qwen3 14B (Ollama) push in the same direction: it is a planner, not a finisher. If your Atlas sessions are mostly mechanical edits, a coder tag will produce cleaner diffs. If your sessions are mostly deciding what to build, Qwen3 14B earns its 9.3GB. Pricing is Free (self-hosted), so the cost of keeping both tags registered is disk, not money. Enable Ollama embeddings for the codebase index so semantic search runs entirely against local vectors, and the whole planning loop stays on the machine. ## Setup 1. Pull the dense mid size: ollama pull qwen3:14b (9.3GB). 2. Expect roughly 11GB of VRAM or unified memory at default context. 3. In atlas.json, under the ollama provider models map, add qwen3:14b with limit.context 40960 and limit.output 8192. 4. Assign it as the model for planning sessions and keep a coder tag registered alongside it. 5. Use /models in the Atlas TUI to switch mid-session between the planner and the coder tag. 6. Enable Ollama embeddings for the codebase index so semantic search runs entirely against local vectors. ## FAQ ### qwen3 14b vs qwen2.5-coder 14b for atlas Qwen3 14B (Ollama) is better at prose, planning, and mixed reasoning; Qwen2.5-Coder 14B is usually better on pure diff generation at the same size. Both are roughly 11GB to serve, so you can register both and switch with /models in the Atlas TUI. ### what context window does qwen3 14b have Qwen3 14B (Ollama) has 40K tokens (40,960), wider than the 32K of the Qwen2.5-Coder line. Register it in the atlas.json ollama models map with limit.context 40960 and limit.output 8192. ### how much vram does qwen3 14b need Expect roughly 11GB of VRAM or unified memory at default context for Qwen3 14B (Ollama). The download itself is 9.3GB via ollama pull qwen3:14b, and the KV cache sits on top of the weights. ### is qwen3 14b dense or mixture of experts Qwen3 14B (Ollama) is dense, with no MoE routing, which makes throughput predictable on a single consumer GPU. The qwen3 MoE tags are a different line and advertise a 256K window, well beyond the 40K tokens (40,960) of the dense 14B. ### how do i switch models mid-session in atlas Run /models in the Atlas TUI. Atlas lets you switch the active model and provider on the fly with favorites and recents, so you can plan on Qwen3 14B (Ollama) and then hand execution to a registered coder tag without restarting. ### is qwen3 14b free to run locally Yes. Qwen3 14B (Ollama) is Free (self-hosted). The costs are the 9.3GB download, roughly 11GB of memory to serve it, and electricity. There is no per-token charge on any Atlas session. ### which local model is best for architecture planning Qwen3 14B (Ollama) suits architecture work rather than pure edits: it is better at prose, planning, and mixed reasoning, and its 40K tokens (40,960) window holds a longer Atlas plan agent transcript than the 32K Qwen2.5-Coder line. --- Canonical HTML: https://runatlas.sh/resources/models/qwen3-14b-local Source of truth: aeo_pages row `/resources/models/qwen3-14b-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.