Qwen3.6 35B-A3B is a 35B total, 3B active mixture-of-experts model from Alibaba and the cheapest reasoning model in the entire Qwen3.6 generation at $0.248 per Mtok input and $1.485 per Mtok output. The odd pricing precision is real. Inside Atlas, Qwen3.6 35B-A3B is the obvious default for the small_model slot on a Qwen-only stack, and it keeps the full 256K tokens (262,144) context window.
Is Qwen3.6 35B-A3B cheap enough to leave running all day?
Qwen3.6 35B-A3B bills $0.248 per Mtok input and $1.485 per Mtok output, less than half the price of the dense Qwen3.6 27B. At that rate an Atlas terminal can stay open through a full working day of tool calls without the invoice becoming a reason to close it.
Atlas is extensible through plugins that contribute tools and hook into agent lifecycle events, and Atlas connects to Model Context Protocol servers and exposes their tools to the agent. Every one of those tools produces a call, a response, and another turn of billing. On an expensive tier that tool traffic is a reason to keep the agent on a short leash. On Qwen3.6 35B-A3B at $1.485 per Mtok output it is not. Atlas reads git branches, status, and diffs, and can stage and create commits on your behalf, so a long-running Qwen3.6 35B-A3B session can also handle the commit hygiene at the end of the work rather than you dropping to a shell to do it.
What does 35B total with 3B active mean for token cost?
Qwen3.6 35B-A3B carries 35B total parameters with roughly 3B active per token, so the compute cost of a token is closer to a small model than a mid-size one. Alibaba passes that through as $0.248 per Mtok input, the cheapest reasoning tier of the Qwen3.6 generation.
The router inside Qwen3.6 35B-A3B fires roughly 3B of the 35B parameters for any given token. Serving cost tracks the active count, not the stored count, which is the entire reason a reasoning-capable model reaches $0.248 per Mtok input in 2026. What matters for Atlas is that the discount does not come with a capability class demotion: Qwen3.6 35B-A3B still reasons, so work handed to it does not collapse into shallow pattern completion. Atlas fans out work to subagents that can run in the foreground or in parallel background sessions, and parallel background subagents are precisely the traffic you want billed at $1.485 per Mtok output rather than at a frontier rate.
Does the cheapest Qwen3.6 tier get a smaller context window?
No. Qwen3.6 35B-A3B still gets the full 262,144 token context and 65,536 token output of the generation, not a cut-down window. Most model families segment the cheap tier by shrinking the window, and Alibaba did not do that here.
Atlas indexes code by AST declarations using tree-sitter, not blind line windows, so the 262,144 tokens Qwen3.6 35B-A3B accepts arrive as complete functions, types, and classes rather than as arbitrary slices. The 65,536 token output ceiling matches the expensive tiers too, which means a single Qwen3.6 35B-A3B turn can emit a substantial unified diff. Atlas computes a unified diff for every file edit and surfaces it for approval before writing, and the review step is identical whether the diff came from a $0.248 tier or a $7.80 one. Breadth, in short, is not what you gave up by choosing Qwen3.6 35B-A3B. Depth is, and that is the next question.
Where does Qwen3.6 35B-A3B fall short?
Qwen3.6 35B-A3B has two documented weaknesses. Roughly 3B active parameters limits its depth on genuinely hard cross-file reasoning, and MoE routing means tail latency is worse than a dense model of the same throughput.
Depth first. A problem whose solution requires holding several coupled subsystems in mind at once is where 3B active parameters runs out, and no amount of the 262,144 token window fixes that: Qwen3.6 35B-A3B can be shown the code without being able to reason across all of it. Atlas drafts a plan in a read-only plan agent and asks before switching to a build agent, so a shallow plan is visible before it becomes a bad edit. Latency second. Routing in Qwen3.6 35B-A3B produces occasional slow turns that a dense model would not, so if you sit watching the Atlas TUI, rendered with SolidJS through the OpenTUI renderer, you will feel those spikes. Neither weakness is a surprise; both are the price of $0.248 per Mtok input.
How do you pin Qwen3.6 35B-A3B in atlas.json?
Export DASHSCOPE_API_KEY in your shell, confirm registry resolution with `atlas models alibaba`, then use Qwen3.6 35B-A3B as the cheap slot: `"small_model": "alibaba/qwen3.6-35b-a3b"` in atlas.json. Two config decisions, one credential, done.
The other option for Qwen3.6 35B-A3B is the main slot. Pin it as the main `"model"` for high-volume mechanical work where $1.485 per Mtok output matters more than depth, such as a long grind of repetitive edits where the marginal value of a deeper model is close to zero. Running Qwen3.6 35B-A3B as the main model does not weaken any Atlas guardrail: every Atlas tool call is permission-gated against allow, ask, and deny rules before it runs, and Atlas snapshots file changes as git patches so edits can be diffed and rolled back. Atlas lets you switch the active model and provider on the fly with favorites and recents, so promoting Qwen3.6 35B-A3B from the small slot to the main one for an afternoon is a keystroke.
Setup
- 01Export DASHSCOPE_API_KEY in your shell.
- 02Confirm registry resolution with `atlas models alibaba`.
- 03Use it as the cheap slot: `"small_model": "alibaba/qwen3.6-35b-a3b"` in atlas.json.
- 04Or pin it as the main `"model"` for high-volume mechanical work where $1.485 per Mtok output matters more than depth.
- 05Keep the permission rules on ask for bash so a cheap high-volume model still asks before it runs shell commands.
Frequently asked questions
- what is the cheapest model in the qwen3.6 family
- Qwen3.6 35B-A3B is the cheapest reasoning model in the entire Qwen3.6 generation at $0.248 per Mtok input and $1.485 per Mtok output, and it still carries the full 262,144 token context.
- how much does qwen3.6 35b-a3b cost
- Qwen3.6 35B-A3B costs $0.248 per Mtok input and $1.485 per Mtok output. The odd pricing precision is real, and it is less than half the price of the dense Qwen3.6 27B.
- does qwen3.6 35b-a3b have the full context window
- Yes. Qwen3.6 35B-A3B gets the full 262,144 token context and 65,536 token output of the Qwen3.6 generation, not a cut-down window.
- what should i set as small_model in atlas
- On a Qwen-only stack, set `"small_model": "alibaba/qwen3.6-35b-a3b"` in atlas.json. At $0.248 per Mtok input, background subagent traffic and session titles cost almost nothing.
- is qwen3.6 35b-a3b good enough as a main coding model
- For high-volume mechanical work, yes. Pin Qwen3.6 35B-A3B as the main `"model"` when $1.485 per Mtok output matters more than depth. Roughly 3B active parameters limits it on genuinely hard cross-file reasoning.
- why is qwen3.6 35b-a3b latency inconsistent
- MoE routing. Qwen3.6 35B-A3B fires about 3B of its 35B parameters per token through a router, and that routing means tail latency is worse than a dense model of the same throughput.
- how do i set up qwen3.6 35b-a3b in atlas
- Export DASHSCOPE_API_KEY in your shell, confirm registry resolution with `atlas models alibaba`, then set it as either `"small_model": "alibaba/qwen3.6-35b-a3b"` or the main `"model"` in atlas.json.
Try Atlas in your terminal
The terminal-native AI coding agent. Free core, single binary.
Install AtlasRelated guides
Research a Third-Party API Before Integrating It with Atlas in 2026
How to research a third-party API with Atlas in 2026: websearch finds the current docs, webfetch pulls the page as markdown or text, and grep checks repo conventions.
Atlas vs Base44: Terminal AI Coding Agents in 2026
Compare Atlas, the terminal-native AI coding agent, with Base44, the Wix-owned no-code app builder, for developers in 2026. Evaluate features, pricing, and workflow.
Atlas vs Traycer in 2026: Terminal Agent That Writes Code vs a Planning Layer Above One
Atlas vs Traycer in 2026. Traycer plans and verifies but writes no code itself, so you pay it on top of an agent. Atlas plans and writes, free and open source.
Review a Pull Request with Atlas (2026 Workflow)
How to review a pull request with Atlas in 2026: bash produces the raw patch, read pulls whole files, the lsp tool's findReferences checks callers the diff never shows.
Atlas vs Crush: Terminal AI Coding Agents in 2026
Comparing Atlas and Crush, two terminal AI coding agents for developers in 2026. Atlas offers robust planning and diffing, while Crush features LSP integration and mid-session model switching.
Atlas vs Jules: Terminal AI Coding Agents in 2026
Atlas and Jules comparison for 2026. Atlas offers terminal-native TUI and local code indexing. Jules provides cloud VMs and a Critic agent for PR review.
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.
Automate GitHub Issue and Pull Request Triage with Atlas (2026 Workflow)
How to automate GitHub issue and pull request triage with Atlas in 2026: the atlas github command checks the actor has admin or write permission before it does anything.