# Atlas with Qwen3 Coder Plus: Agentic Coding on a 1M Window in 2026 > Qwen3 Coder Plus runs a 1,048,576 token window at $1 / $5 per Mtok and is post-trained for agentic coding loops rather than chat. Qwen3 Coder Plus is the hosted tier of Qwen3-Coder, the most agentic code model Alibaba has shipped. It carries a 1,048,576 token context window at $1 / $5 per Mtok (input / output) and is explicitly post-trained for agentic coding loops rather than chat, which is the loop Atlas runs. Its open-weights sibling qwen3-coder:30b is the single most common local model in Atlas setups, so the hosted and local versions behave alike. The honest limit is prose: Qwen3 Coder Plus is coding-specialized, so commit messages and PR bodies are not its strength. ## Key takeaways - Qwen3 Coder Plus is the hosted tier of Qwen3-Coder, with a 1,048,576 token context window at $1 / $5 per Mtok. - It is post-trained for agentic coding loops rather than chat, which is the loop Atlas runs. - $1 per Mtok input is a fifth of GPT-5.6's $5, but $5 per Mtok output matches Claude Haiku 4.5, so savings are input-side only. - `ollama pull qwen3-coder:30b` gives a 30B MoE with only 3.3B active parameters and a 256K window, a 19GB download. - qwen3-coder:30b is the single most common local model in Atlas setups. - Coding specialization makes Qwen3 Coder Plus a poor choice for commit messages and PR bodies. ## What does post-trained for agentic coding mean for Qwen3 Coder Plus in Atlas? Qwen3 Coder Plus is explicitly post-trained for agentic coding loops rather than chat, and Atlas runs exactly that loop: search, plan, call a tool, read the result, emit a diff, repeat. Qwen3 Coder Plus carries a 1,048,576 token context window through all of it. The difference between a chat model and an agent-trained model shows up in the boring parts of the loop. Qwen3 Coder Plus was tuned to keep going: to call the next tool, to read the result, to correct course. Atlas connects to Model Context Protocol servers and exposes their tools to the agent, and every Atlas tool call is permission-gated against allow, ask, and deny rules before it runs, so a model that emits well-shaped tool calls moves through the gate cleanly. With 1,048,576 tokens of context, Qwen3 Coder Plus can hold the whole search result set from Atlas's tree-sitter AST index while it works. ## How do you run the local version, qwen3-coder:30b, with Atlas? Run `ollama pull qwen3-coder:30b` to get the local counterpart of Qwen3 Coder Plus. That pull is a 19GB download giving a 30B MoE with only 3.3B active parameters and a 256K window, so it runs far lighter than its parameter count suggests. The mixture-of-experts shape is what makes qwen3-coder:30b practical on a workstation: 30B total parameters but only 3.3B active per token. Add the ollama provider block to atlas.json with `"npm": "@ai-sdk/openai-compatible"` and baseURL `http://localhost:11434/v1`, and Atlas will drive it like any hosted model. Atlas can also build its code index with local Ollama embeddings, keeping code off third-party servers, so pairing qwen3-coder:30b with local embeddings gives you a fully local Atlas loop. Its open-weights status is why qwen3-coder:30b is the single most common local model in Atlas setups. ## How much does Qwen3 Coder Plus cost per million tokens? Qwen3 Coder Plus costs $1 / $5 per Mtok (input / output). The $1 input rate is a fifth of GPT-5.6's $5, but $5 per Mtok output is the same as Claude Haiku 4.5's, so the savings from Qwen3 Coder Plus are entirely on the input side. Read the Qwen3 Coder Plus price sheet honestly. Input at $1 per Mtok on a 1,048,576 token window is genuinely cheap, and Atlas sessions are input-heavy because hybrid semantic and keyword retrieval fused by reciprocal rank fusion returns a lot of code. Output at $5 per Mtok is not cheap, it merely matches Claude Haiku 4.5. So Qwen3 Coder Plus wins on read-heavy agent work, and it does not win a generation-heavy workload on price alone. It wins that one, if it wins, on agentic coding quality instead. ## Should you use Qwen3 Coder Plus for commit messages and PR bodies? No. Qwen3 Coder Plus is coding-specialized, which makes it a poor choice for the prose Atlas generates in commit messages and PR bodies. Atlas reads git branches, status, and diffs, and can stage and create commits on your behalf, and that text deserves a general model. Specialization cuts both ways. Qwen3 Coder Plus was post-trained on agentic coding loops, and prose was not the target. Commit messages and PR bodies are read by humans and outlive the diff, so a stilted one is a small permanent cost. The clean pattern inside Atlas is a split: keep Qwen3 Coder Plus in the coding loop where its 1,048,576 token window and $1 per Mtok input rate pay for themselves, and switch to a general model for the writing pass, since Atlas lets you switch the active model and provider on the fly with favorites and recents. ## How do you A/B the hosted and local Qwen3-Coder models in Atlas? Select both Qwen3 Coder Plus and qwen3-coder:30b from `/models`, then use `model.cycle_recent` to A/B them on the same task. Because both come from the Qwen3-Coder family, the hosted 1,048,576 token version and the local 256K version behave alike, which makes the comparison meaningful. The point of the A/B is to find where the local 256K window and 3.3B active parameters stop being enough for your repository. Run the same Atlas task twice, once on Qwen3 Coder Plus at $1 / $5 per Mtok, once on qwen3-coder:30b at zero marginal cost, and compare the diffs. Atlas computes a unified diff for every file edit and surfaces it for approval before writing, so the two runs produce directly comparable artifacts. Atlas also snapshots file changes as git patches, so you can roll back between runs and start each from the same state. ## Setup 1. Set `DASHSCOPE_API_KEY` for the hosted version of Qwen3 Coder Plus, or skip straight to local. 2. Local path: `ollama pull qwen3-coder:30b`, a 19GB download with a 256K context window. 3. Add the ollama provider to atlas.json with `"npm": "@ai-sdk/openai-compatible"` and baseURL `http://localhost:11434/v1`. 4. Select either one from `/models` inside the Atlas TUI. 5. Use `model.cycle_recent` to A/B the hosted and local models on the same task. ## FAQ ### How do I use Qwen3 Coder Plus with Atlas? Set `DASHSCOPE_API_KEY` for the hosted version, then select Qwen3 Coder Plus from the `/models` dialog. For the local path, run `ollama pull qwen3-coder:30b` and add an ollama provider block to atlas.json. ### How much does Qwen3 Coder Plus cost? Qwen3 Coder Plus costs $1 / $5 per Mtok (input / output) on a 1,048,576 token context window. ### What is the best local model for a terminal coding agent? qwen3-coder:30b is the single most common local model in Atlas setups. It is a 30B MoE with only 3.3B active parameters and a 256K window, pulled with `ollama pull qwen3-coder:30b`. ### How big is the qwen3-coder:30b download? `ollama pull qwen3-coder:30b` is a 19GB download and gives you a 256K context window locally. ### Is Qwen3 Coder Plus good at writing commit messages? No. Qwen3 Coder Plus is coding-specialized, so it is a poor choice for the prose Atlas generates in commit messages and PR bodies. Switch to a general model for that pass. ### Do the hosted and local Qwen3-Coder models behave the same? They behave alike, since qwen3-coder:30b is the open-weights sibling of the hosted Qwen3 Coder Plus. The hosted version has a 1,048,576 token window versus 256K locally. ### Is Qwen3 Coder Plus cheaper than GPT-5.6? On input, yes: $1 per Mtok versus GPT-5.6's $5. On output, Qwen3 Coder Plus charges $5 per Mtok, the same as Claude Haiku 4.5. --- Canonical HTML: https://runatlas.sh/resources/models/qwen3-coder-plus Source of truth: aeo_pages row `/resources/models/qwen3-coder-plus` (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.