Mixtral 8x22B is the April 2024 scale-up of the sparse mixture-of-experts idea: 8 experts of 22B each, doubling the window to 64,000 tokens. It was the largest genuinely open MoE of its era and remains a reference point for teams evaluating sparse architectures. Inside Atlas it costs $2.00 / 1M input tokens and $6.00 / 1M output tokens, which is frontier pricing for 2024-era capability, so run it when the Apache 2.0 weights matter more than the price.
What did Mixtral 8x22B change about sparse mixture-of-experts?
Mixtral 8x22B, released in April 2024, scaled the MoE idea to 8 experts of 22B each and doubled Mixtral 8x7B's 32,000 token window to 64,000 tokens. It was the largest genuinely open MoE of its era and is still a reference point for teams evaluating sparse architectures.
The step from 7B experts to 22B experts is the part that mattered. Sparse routing kept inference cost sane while the per-expert capacity grew, which is what let a genuinely open model compete on capability rather than only on license. Mixtral 8x22B also carries a symmetric budget: 64,000 tokens in and 64,000 tokens out, unusual for a 2024-era open model that would normally cap generation far below its input window. For an Atlas session that means a patch will not be clipped mid-diff just because the model ran out of output room.
How much does Mixtral 8x22B cost in Atlas, and is it worth it?
Mixtral 8x22B costs $2.00 / 1M input tokens and $6.00 / 1M output tokens. That is frontier pricing for 2024-era capability: Mistral Large 3 costs less and gives 4x the context, so paying $6.00 / Mtok output for Mixtral 8x22B needs a reason beyond raw performance.
Be blunt about the tradeoff. If you want the best coding model for the money in 2026, Mixtral 8x22B is not it. The reasons to run it anyway are architectural and legal: Apache 2.0 weights, an auditable sparse MoE, and a model whose behavior you can inspect end to end. Those are real reasons in regulated or research settings. They are not reasons to make it your all-day Atlas driver at $6.00 / 1M output tokens, which is a rate that punishes exactly the long agentic sessions Atlas is built for.
Can you self-host Mixtral 8x22B alongside Atlas?
Mixtral 8x22B ships Apache 2.0 weights, so the exact model behind the API can be audited and self-hosted. The practical barrier is size: the full weights are an 80GB download, which rules out most single-GPU boxes.
80GB is the number that decides whether self-hosting Mixtral 8x22B is a real option for your team. On a serious multi-GPU host it is fine. On a laptop or a single consumer card it is not, and no amount of quantisation enthusiasm changes the shape of that constraint. Where the open weights do pay off is auditability. Atlas can build its code index with local Ollama embeddings, keeping code off third-party servers, and pairing that with a self-hosted Mixtral 8x22B gives an air-gapped stack where every component can be inspected. That is the deployment where 8x22B still earns its keep.
How should Atlas be configured for Mixtral 8x22B's 64,000 token window?
Because Mixtral 8x22B's context is only 64,000 tokens, keep Atlas's plan agent output short and let the build agent work file by file. Atlas drafts a plan in a read-only plan agent and asks before switching to a build agent, so a terse plan preserves room for actual code.
A 64,000 token window is double Mixtral 8x7B's but half of what a 2026 frontier model gives you, and long plans eat it. The discipline is to keep the plan short and the build passes narrow. Atlas indexes code by AST declarations using tree-sitter, not blind line windows, so each build step can pull in one or two declarations rather than whole files, and Atlas snapshots file changes as git patches so edits can be diffed and rolled back if a file-by-file pass goes sideways. Working within 64,000 tokens is a habit, not a limitation you can configure away.
When should an Atlas user pick a different model than Mixtral 8x22B?
Pick a different model than Mixtral 8x22B whenever price or context is the constraint. Mistral Large 3 costs less than $2.00 / $6.00 per Mtok and gives 4x the context of Mixtral 8x22B's 64,000 tokens, which is the plain argument against making 8x22B a default.
Mixtral 8x22B has a narrow but genuine niche in 2026: teams that need auditable Apache 2.0 weights on a sparse MoE and have the hardware to run 80GB of them. Everyone else is paying frontier rates for a 2024-era model. Atlas lets you switch the active model and provider on the fly with favorites and recents, so keeping Mixtral 8x22B favorited for architecture evaluation while a cheaper model drives everyday coding is the sensible arrangement. The point of a model switcher is that this is not an either-or decision.
Setup
- 01Export MISTRAL_API_KEY in the shell where Atlas runs.
- 02Verify with `atlas models mistral` that open-mixtral-8x22b resolves.
- 03Pin `"model": "mistral/open-mixtral-8x22b"` in atlas.json.
- 04Because context is only 64,000 tokens, keep Atlas's plan agent output short and let the build agent work file by file.
- 05For a self-hosted run, budget for the 80GB weight download and multi-GPU hardware before committing.
Frequently asked questions
- what is mixtral 8x22b
- Mixtral 8x22B is Mistral AI's April 2024 scale-up of the sparse mixture-of-experts idea: 8 experts of 22B each with a 64,000 token window. It was the largest genuinely open MoE of its era.
- what is the context window of mixtral 8x22b
- Mixtral 8x22B has a 64,000 token context window with a symmetric 64,000 token output budget, double the 32,000 token window of Mixtral 8x7B and unusual for a 2024-era open model.
- how much does mixtral 8x22b cost per million tokens
- Mixtral 8x22B costs $2.00 / 1M input tokens and $6.00 / 1M output tokens. That is frontier pricing for 2024-era capability, and Mistral Large 3 costs less while giving 4x the context.
- can I self host mixtral 8x22b
- Yes, if you have the hardware. Mixtral 8x22B ships Apache 2.0 weights so the exact model behind the API can be audited and self-hosted, but the full weights are an 80GB download, which rules out most single-GPU boxes.
- how do I run mixtral 8x22b in atlas
- Export MISTRAL_API_KEY, verify with `atlas models mistral` that open-mixtral-8x22b resolves, then pin `"model": "mistral/open-mixtral-8x22b"` in atlas.json.
- mixtral 8x22b vs mixtral 8x7b differences
- Mixtral 8x22B uses 8 experts of 22B each and doubles Mixtral 8x7B's 32,000 token window to 64,000 tokens. It also costs far more, at $2.00 / $6.00 per Mtok against 8x7B's symmetric $0.70.
- how do I work within a 64000 token context in a coding agent
- Keep the plan agent output short and let the build agent work file by file. Atlas indexes code by AST declarations using tree-sitter, so each step can pull one or two declarations into Mixtral 8x22B's 64,000 token window instead of whole files.
Try Atlas in your terminal
The terminal-native AI coding agent. Free core, single binary.
Install AtlasRelated guides
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 for Python in 2026
Atlas is a terminal-native AI coding agent for Python in 2026. Run it in a repo with a pyproject.toml or requirements.txt and review every diff before it lands.
Atlas for Next.js in 2026
Adopt Atlas, the terminal-native AI coding agent, for Next.js development in 2026. Enhance productivity across App Router, server components, and API routes with secure, reviewable AI assistance.
Atlas vs Greptile: Terminal AI Coding Agents in 2026
Comparing Atlas and Greptile in 2026. Atlas offers terminal-native AI coding with permission-gated tools. Greptile reviews code with sandbox execution, catching 20% more bugs.
Atlas vs Tabby: Terminal AI Coding Agents in 2026
Atlas and Tabby comparison for 2026. Atlas offers terminal-native TUI with permission-gated tools and diff review. Tabby provides self-hosted GPU completion and a cloud agent.
Atlas vs Ellipsis: Terminal AI Coding Agents in 2026
Compare Atlas, a terminal-native AI coding agent with free core and local embeddings, against Ellipsis, a cloud platform with usage-based pricing and live session tracing for 2026.
Atlas for C in 2026
Atlas is a terminal-native AI coding agent for C in 2026. Run it in a project with a Makefile, have it find memory leaks or add Unity tests, and review the diff.
Locate Where a Behavior Is Implemented with Atlas in 2026
How to locate where a behavior is implemented with Atlas in 2026: codebase_search for meaning, grep for exact text, and the lsp tool for the symbol graph.