# Atlas with Qwen3.5 397B-A17B: The Qwen3.5 Flagship in 2026 > Qwen3.5 397B-A17B delivers frontier-scale reasoning at $0.60 per Mtok input, with 262,144 tokens of context and 65,536 tokens of output. Qwen3.5 397B-A17B is the flagship mixture-of-experts model of Alibaba's Qwen3.5 generation, with 397B total parameters and 17B active per token, and it is the top open-architecture reasoning tier Alibaba lists for this line. In Atlas it is the model for repository-scale planning and deep debugging. Qwen3.5 397B-A17B carries a 256K tokens (262,144) context and prices at $0.60 per Mtok input and $3.60 per Mtok output, below most Western frontier models on the input side. ## Key takeaways - Qwen3.5 397B-A17B has 397B total parameters with 17B active per token, the largest Qwen3.5 tier and the deepest reasoning in the family. - $0.60 per Mtok input buys frontier-scale reasoning, an order of magnitude below comparable Western flagships. - 262,144 tokens of context plus 65,536 tokens of output is enough for Atlas to plan and then emit a repo-wide diff. - $3.60 per Mtok output adds up fast in an agentic loop that generates diffs on every turn. - 17B active parameters is fewer than Qwen3-Coder 480B-A35B's 35B, so on pure coding agentics the Coder line can still win. ## What is Qwen3.5 397B-A17B best at inside Atlas? Qwen3.5 397B-A17B is the deepest reasoning tier in the Qwen3.5 family, with 397B total parameters and 17B active per token. Inside Atlas it earns its $0.60 per Mtok input on the hardest work: repository-scale planning, subtle bugs, and architectural decisions the cheaper tiers get confidently wrong. The work that justifies Qwen3.5 397B-A17B is the work where being wrong is expensive. Atlas drafts a plan in a read-only plan agent and asks before switching to a build agent, and running that plan phase on Qwen3.5 397B-A17B means the highest-capability model in the line is the one deciding what the change should be, before any file is touched. Atlas searches code with hybrid semantic and keyword retrieval fused by reciprocal rank fusion and indexes code by AST declarations using tree-sitter, so what Qwen3.5 397B-A17B is reasoning over is real, whole declarations pulled from across the repository. 262,144 tokens of context plus 65,536 tokens of output is enough for Atlas to plan and then emit a repo-wide diff in one session. ## How does Qwen3.5 397B-A17B pricing compare to Western frontier models? Qwen3.5 397B-A17B charges $0.60 per Mtok input for frontier-scale reasoning, an order of magnitude below comparable Western flagships. The output side at $3.60 per Mtok is the number to watch, because it adds up fast in an agentic loop that generates diffs on every turn. Input pricing is where Qwen3.5 397B-A17B is genuinely striking. $0.60 per Mtok for a 397B-parameter reasoning model with a 262,144 token window means the retrieval-heavy half of an Atlas session, which is the half with the most tokens, is close to free relative to Western frontier tiers. The honest counterweight is output. $3.60 per Mtok output adds up fast in an agentic loop that generates diffs on every turn, and Atlas is exactly such a loop: every accepted edit is output tokens, every plan is output tokens, every reasoning trace is output tokens. The discipline that keeps Qwen3.5 397B-A17B affordable is to concentrate its output on decisions, not on volume, and to let a cheaper tier carry the mechanical edits. ## Where does Qwen3.5 397B-A17B lose to Qwen3-Coder? Qwen3.5 397B-A17B activates 17B parameters per token, which is fewer than Qwen3-Coder 480B-A35B's 35B active. On pure coding agentics the Coder line can still win, and pretending otherwise would not help anyone choosing a model for Atlas in 2026. Active parameters are the honest yardstick for a mixture-of-experts model, and Qwen3-Coder 480B-A35B fires more than twice as many per token as Qwen3.5 397B-A17B. On tasks that are specifically coding agentics, long tool-use chains, precise multi-file edits, sticking to a plan across many turns, that gap shows. Qwen3.5 397B-A17B is the deepest general reasoner in the Qwen3.5 family, and it plans beautifully, but if your Atlas usage is mostly the build agent grinding through edits rather than the plan agent thinking, the Coder line deserves a real trial. Atlas lets you switch the active model and provider on the fly with favorites and recents, so running the plan on Qwen3.5 397B-A17B and the build on a Coder tier is one keystroke, not a config rewrite. ## How do you run Atlas safely on Qwen3.5 397B-A17B? Run the plan agent first, review the unified diff, and keep bash on ask rather than allow for the first few sessions. Qwen3.5 397B-A17B has 65,536 tokens of output and will happily emit a repo-wide diff, so the guardrails matter more at flagship scale, not less. Safety practice with Qwen3.5 397B-A17B in Atlas is about capping blast radius while the model is producing changes at flagship scale. Every Atlas tool call is permission-gated against allow, ask, and deny rules before it runs, so keeping bash on ask for the first few sessions means you see each shell command the model wants to run before it runs. Atlas computes a unified diff for every file edit and surfaces it for approval before writing, and Atlas snapshots file changes as git patches so edits can be diffed and rolled back. Those three together mean a bad plan from Qwen3.5 397B-A17B is caught at review, not at merge. Atlas reads git branches, status, and diffs, and can stage and create commits on your behalf, so the model can also do the cleanup work once you approve. ## Is the 262,144 token window enough for repository-scale work? Qwen3.5 397B-A17B carries 262,144 tokens of context plus 65,536 tokens of output, enough for Atlas to plan and then emit a repo-wide diff. Whether that is enough depends on retrieval quality, not on the raw number, and Atlas retrieves rather than dumps. A 262,144 token window sounds like a hard limit on how much repository Qwen3.5 397B-A17B can consider, but in Atlas the operative constraint is what gets selected into that window. Atlas indexes code by AST declarations using tree-sitter, not blind line windows, and fuses semantic and keyword retrieval with reciprocal rank fusion, so the 262,144 tokens fill with relevant declarations rather than adjacent noise. That makes the window go far further than a naive file dump would. If you genuinely need a larger raw window, the hosted Plus tiers in the Qwen line reach 1,000,000 tokens, at the cost of closed weights. For most repository-scale planning with Qwen3.5 397B-A17B, 262,144 tokens plus good retrieval is the working answer. ## Setup 1. Set DASHSCOPE_API_KEY, or store the key with `atlas login` under the Alibaba provider. 2. Confirm with `atlas models alibaba` that `qwen3.5-397b-a17b` resolves. 3. Pin it: `"model": "alibaba/qwen3.5-397b-a17b"` in atlas.json. 4. Run the plan agent first, review the unified diff, and keep bash on ask rather than allow for the first few sessions. 5. Let Atlas's tree-sitter AST indexing and RRF-fused search fill the 262,144 token window instead of pasting files. ## FAQ ### how much does qwen3.5 397b-a17b cost Qwen3.5 397B-A17B from Alibaba costs $0.60 per Mtok input and $3.60 per Mtok output. The $0.60 input price is an order of magnitude below comparable Western flagships. ### what is the best qwen model for a coding agent Qwen3.5 397B-A17B is the deepest reasoning tier in the Qwen3.5 family, but its 17B active parameters is fewer than Qwen3-Coder 480B-A35B's 35B, so on pure coding agentics the Coder line can still win. Plan on the 397B, build on a Coder tier. ### what is the context window of qwen3.5 397b-a17b Qwen3.5 397B-A17B has a 256K tokens (262,144) context window plus 65,536 tokens of output, enough for Atlas to plan and then emit a repo-wide diff. ### how do i set up qwen3.5 397b-a17b in atlas Set DASHSCOPE_API_KEY or store the key with `atlas login` under the Alibaba provider, confirm with `atlas models alibaba` that `qwen3.5-397b-a17b` resolves, then pin `"model": "alibaba/qwen3.5-397b-a17b"` in atlas.json. ### how do i control costs on a frontier model in an agent loop Watch output, not input. Qwen3.5 397B-A17B is $0.60 per Mtok input but $3.60 per Mtok output, and an agentic loop generates diffs on every turn. Concentrate the flagship on planning and let a cheaper tier carry mechanical edits. ### what does a17b mean in qwen3.5 397b-a17b A17B means Qwen3.5 397B-A17B activates 17B of its 397B total parameters per token. It is a mixture-of-experts model, and the active count is what predicts reasoning depth per turn. ### should i keep bash on ask when running a frontier model in atlas Yes, for the first few sessions with Qwen3.5 397B-A17B. Every Atlas tool call is permission-gated against allow, ask, and deny rules before it runs, and keeping bash on ask lets you see each shell command before a 65,536 token output run executes it. --- Canonical HTML: https://runatlas.sh/resources/models/qwen3-5-397b-a17b Source of truth: aeo_pages row `/resources/models/qwen3-5-397b-a17b` (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.