# Atlas with DeepCoder 14B (Ollama): RL-Tuned for First-Attempt Diffs in 2026 > DeepCoder 14B (Ollama) is a 9.0GB RL-tuned coder with a 128K token (131,072) context, Free (self-hosted), and it targets first-attempt correctness. DeepCoder 14B (Ollama) is a reinforcement-learning-tuned coder from Agentica and Together AI, built for the pass-at-one case: get it right on the first attempt. It is a 9.0GB download with a context window of 128K tokens (131,072), which is a strong pairing for an Atlas agent that must produce a correct diff, not a plausible one. Pricing is Free (self-hosted). The weights are published as preview, so treat behavior as less settled than a mainline release. ## Key takeaways - DeepCoder 14B (Ollama) is a 9.0GB download with a 128K token (131,072) context, four times the window of Qwen2.5-Coder 14B at nearly identical size. - RL post-training from Agentica and Together AI targets first-attempt correctness, the metric that matters when Atlas applies a unified diff. - Roughly 11GB to serve, so it competes for the same 12GB card as the mainstream 14B coders. - Pricing is Free (self-hosted), so long runs against the 128K window carry no per-token cost. - The weights are published as preview, so treat behavior as less settled than a mainline Qwen or DeepSeek release. ## Why does first-attempt correctness matter for an Atlas agent? DeepCoder 14B (Ollama) received RL post-training aimed specifically at first-attempt correctness, which is exactly the metric that matters when Atlas is applying a unified diff. Atlas computes a diff for every file edit and surfaces it for approval, so a plausible patch that fails is wasted review time. Most code models are optimized for output that looks right. An agent needs output that is right, because the human in the loop is reviewing a unified diff and approving a write to disk. DeepCoder 14B from Agentica and Together AI was reinforcement-learning-tuned for pass-at-one, the case where the first attempt has to land. Inside Atlas, use the unified-diff review to check first-attempt correctness before every write, and you will find the review cheaper: fewer rejected patches, fewer regeneration rounds. That is the specific claim of this model, and it is the reason to pick it over a general 14B coder. ## How does DeepCoder 14B compare to other 14B coders on context? DeepCoder 14B (Ollama) carries a 128K token (131,072) context at a 9.0GB download, four times the window of Qwen2.5-Coder 14B at nearly identical size. For an Atlas session that must hold retrieved code, a plan, and a tool transcript, that ratio is the headline number. Window size at a fixed footprint is the cleanest comparison you can make between models in the same tier, and DeepCoder 14B wins it. Four times the context of Qwen2.5-Coder 14B, from the same 9.0GB of disk, means Atlas compacts far less often during a long run. Register deepcoder:14b in the atlas.json ollama provider with limit.context 131072 and limit.output 8192 to actually use that window. Keep embeddings on Ollama so the 128K window can be filled from a locally built index with no cloud round trip. ## How much VRAM does DeepCoder 14B need? DeepCoder 14B (Ollama) takes roughly 11GB to serve against a 9.0GB download, so it competes for the same 12GB card as the mainstream 14B coders. Raise Ollama's num_ctx toward 128K only as memory allows: 11GB is the weight floor, not the long-context figure. The gap between the weight floor and the real memory cost is the thing to plan for with DeepCoder 14B. Roughly 11GB gets the model resident. Filling the 128K token (131,072) window grows the KV cache well past that, and a 12GB card will run out. The correct procedure is incremental: set limit.context 131072 in atlas.json so Atlas knows the ceiling, then raise Ollama's num_ctx step by step while watching memory, and stop where your hardware stops. You get the long window when you have the memory for it, not because the config says so. ## What does DeepCoder 14B cost to run with Atlas? DeepCoder 14B (Ollama) is Free (self-hosted). The 9.0GB download and roughly 11GB of serving memory are the whole cost, which is a 12GB consumer card in 2026, and there is no per-token bill for a long Atlas run against the 128K token (131,072) window. Free (self-hosted) and first-attempt correctness compound. A metered model that needs three tries costs three times as much; a self-hosted model that needs one try costs nothing either way, but it costs you far less attention. That is the real saving with DeepCoder 14B: reviewer time, not dollars. Keep embeddings on Ollama so the retrieval half of the loop is local too, and the entire pipeline, index, retrieve, generate, diff, runs on the same 12GB card with no cloud round trip and nothing leaving the machine. ## What are the risks of running DeepCoder 14B? DeepCoder 14B (Ollama) is published as preview weights, so treat its behavior as less settled than a mainline Qwen or DeepSeek release. A 1.5b tag also exists, but at that scale the RL tuning does not compensate for capacity, so the 9.0GB tag is the one to run. Preview status is a real caveat on DeepCoder 14B. Mainline releases have been shaken out across many workloads; a preview has not, and quirks may surface in yours. Mitigate with the controls Atlas already gives you: every tool call is permission-gated against allow, ask, and deny rules before it runs, and every file edit is surfaced as a unified diff for approval before writing. Atlas also snapshots file changes as git patches, so edits can be diffed and rolled back. Run the model, review its diffs, and keep a mainline coder registered as the fallback. ## Setup 1. Pull it: ollama pull deepcoder:14b (9.0GB, aliased to :latest). 2. Register deepcoder:14b in the atlas.json ollama provider with limit.context 131072 and limit.output 8192. 3. Raise Ollama's num_ctx toward 128K only as memory allows; the 11GB figure is the weight floor, not the long-context figure. 4. Set model to ollama/deepcoder:14b and use Atlas unified-diff review to check first-attempt correctness before every write. 5. Keep embeddings on Ollama so the 128K window can be filled from a locally built index with no cloud round trip. ## FAQ ### how to run deepcoder 14b with atlas Run ollama pull deepcoder:14b (9.0GB, aliased to :latest), register deepcoder:14b in the atlas.json ollama provider with limit.context 131072 and limit.output 8192, then set model to ollama/deepcoder:14b. ### what is the context window of deepcoder 14b DeepCoder 14B (Ollama) has a 128K token (131,072) context window, four times the window of Qwen2.5-Coder 14B at nearly identical size. ### how much vram does deepcoder 14b need Roughly 11GB to serve against a 9.0GB download, so it competes for the same 12GB card as the mainstream 14B coders. That 11GB is the weight floor, not the long-context figure. ### what does rl tuning do for a coding model DeepCoder 14B received RL post-training aimed specifically at first-attempt correctness, the pass-at-one case. That is exactly the metric that matters when Atlas is applying a unified diff rather than offering a suggestion. ### is deepcoder 14b production ready DeepCoder 14B is published as preview weights, so treat its behavior as less settled than a mainline Qwen or DeepSeek release. Keep Atlas permission gating on and review every unified diff. ### should i use the deepcoder 1.5b tag A 1.5b tag exists, but at that scale the RL tuning does not compensate for capacity. Run the 9.0GB deepcoder:14b tag instead. ### deepcoder 14b vs qwen2.5-coder 14b DeepCoder 14B offers a 128K token (131,072) context, four times the window of Qwen2.5-Coder 14B at nearly identical size, plus RL tuning for first-attempt correctness. Qwen2.5-Coder 14B is a mainline release rather than preview weights. --- Canonical HTML: https://runatlas.sh/resources/models/deepcoder-14b-local Source of truth: aeo_pages row `/resources/models/deepcoder-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.