MiniMax-M2 is a 229.9B parameter mixture-of-experts model that activates only about 10B parameters per token, and that activation ratio is the trick behind its $0.30 per Mtok input and $1.20 per Mtok output pricing. Inside Atlas, MiniMax-M2 is the model for high-volume agent loops where you want reasoning without watching the meter. MiniMax-M2 gives Atlas a 196,608 token context window, and MiniMax uniquely exposes an Anthropic-compatible API, so Atlas talks to it through @ai-sdk/anthropic.
What makes MiniMax-M2 cheap enough to run Atlas all day?
MiniMax-M2 carries 229.9B total parameters but activates roughly 9.8B per token across 256 experts in a 62-layer decoder. That activation ratio, under 5 percent, is what directly explains the $0.30 per Mtok input price and makes MiniMax-M2 practical for continuous Atlas usage.
Sparse activation changes what you can afford to do with a coding agent. Atlas searches code with hybrid semantic and keyword retrieval fused by reciprocal rank fusion, and on an expensive model you ration that search to keep input tokens down. On MiniMax-M2 at $0.30 per Mtok input, a generous retrieval pass costs almost nothing, so you can let Atlas pull the surrounding declarations rather than hand-picking two files and hoping. Atlas indexes code by AST declarations using tree-sitter, not blind line windows, which means the extra tokens you buy at thirty cents are whole functions and types rather than truncated fragments. The economics and the indexing strategy reinforce each other.
Why does MiniMax-M2 fit Atlas's build-and-verify loop?
MiniMax-M2 uses interleaved thinking with explicit <think> tags and was trained for coding-run-fix loops and test-validated repairs. That training target is precisely Atlas's build-and-verify cycle, which makes MiniMax-M2 a better structural match than its $1.20 per Mtok output price would suggest.
Atlas drafts a plan in a read-only plan agent and asks before switching to a build agent, then computes a unified diff for every file edit and surfaces it for approval before writing. A model trained on run-fix loops behaves well in that shape: it proposes an edit, runs the tests, reads the failure, and repairs rather than restating. Interleaved thinking means the reasoning is threaded through the tool calls instead of front-loaded into one block. One caution follows from the same property: interleaved-thinking models issue tool calls aggressively, so keep Atlas's permission gating on allow, ask, and deny rules while you calibrate MiniMax-M2 on a new repository.
How does Atlas connect to MiniMax's Anthropic-compatible API?
Atlas loads MiniMax through `@ai-sdk/anthropic` against https://api.minimax.io/anthropic/v1, because MiniMax exposes an Anthropic-compatible API rather than an OpenAI-shaped one. Setting `"model": "minimax/MiniMax-M2"` in atlas.json is enough, but note the capitalized model id, which is a common source of a failed lookup.
The compatibility layer is a real convenience and a real caveat at the same time. If you already have Atlas wired for Claude, the MiniMax provider block is a base URL and a key rather than a rewrite. But the Anthropic-compatible endpoint is a compatibility layer, not Anthropic, so beta headers and features Atlas sends to real Claude do not all apply on MiniMax-M2. Treat unexpected behavior around newer streaming or tool-call features as a shim gap rather than a model failure, and verify with `atlas models minimax` that `MiniMax-M2` resolves before you debug anything deeper.
What is the catch with MiniMax-M2's 196,608 token context?
MiniMax-M2's 196,608 token context is the smallest in MiniMax's current lineup. MiniMax-M2.1 through M2.7 all moved to 204,800 tokens, and MiniMax-M3 jumped to 1,000,000, so MiniMax-M2 is the one model in the family carrying a window that its own successors have already outgrown.
In practice the gap between 196,608 and 204,800 is small, roughly 4 percent, and it will rarely be the thing that breaks a session. The reason to care is that the newer checkpoints hold the same $0.30 per Mtok input and $1.20 per Mtok output pricing while offering the larger window, so staying on MiniMax-M2 buys you nothing on cost. If your repository work regularly pushes past 200,000 tokens of context, the answer is not MiniMax-M2 at all: MiniMax-M3 gives you 1,000,000 tokens at the same rates.
Why do MiniMax-M2's open weights matter for Atlas users?
MiniMax-M2 ships open weights on HuggingFace at MiniMaxAI/MiniMax-M2, which means the hosted $0.30 per Mtok input price is a convenience rather than a lock-in. Teams that later need on-premise inference can move the same 229.9B MoE checkpoint without rewriting the agent that sits on top of it.
Open weights change the risk profile of standardizing on a model. If MiniMax raises the price or deprecates the endpoint, the checkpoint at MiniMaxAI/MiniMax-M2 is still there. Atlas reinforces that independence on the retrieval side: Atlas can build its code index with local Ollama embeddings, keeping code off third-party servers, so an Atlas deployment on MiniMax-M2 can already keep the index local while only chat traffic leaves the building. For teams evaluating data residency, that split is the useful lever, because MiniMax's hosted endpoint is China-based.
Setup
- 01Get a MiniMax key, then export MINIMAX_API_KEY or run `atlas login` and select MiniMax.
- 02Run `atlas models minimax` and confirm `MiniMax-M2` resolves. Note the capitalized model id.
- 03Set `"model": "minimax/MiniMax-M2"` in atlas.json. Atlas loads MiniMax through `@ai-sdk/anthropic` against https://api.minimax.io/anthropic/v1.
- 04Keep Atlas's permission gating on allow, ask, and deny rules while you calibrate, since interleaved-thinking models issue tool calls aggressively.
- 05Because input is $0.30 per Mtok, let Atlas's hybrid semantic and keyword retrieval pull generous context rather than hand-picking files.
Frequently asked questions
- how much does minimax-m2 cost per million tokens
- MiniMax-M2 costs $0.30 per Mtok input and $1.20 per Mtok output. The low price comes from a sparse MoE design that activates only about 9.8B of its 229.9B parameters per token.
- what is the minimax-m2 context window
- MiniMax-M2 has a 196,608 token context window. That is the smallest in MiniMax's current lineup, since MiniMax-M2.1 through M2.7 all use 204,800 and MiniMax-M3 uses 1,000,000.
- does minimax-m2 use an openai compatible api
- No. MiniMax exposes an Anthropic-compatible API, so Atlas connects to MiniMax-M2 through @ai-sdk/anthropic against https://api.minimax.io/anthropic/v1.
- how do i configure minimax-m2 in atlas
- Export MINIMAX_API_KEY or run `atlas login` and select MiniMax, run `atlas models minimax` to confirm `MiniMax-M2` resolves, then set `"model": "minimax/MiniMax-M2"` in atlas.json. The model id is capitalized.
- are minimax-m2 weights open source
- MiniMax-M2 ships open weights on HuggingFace at MiniMaxAI/MiniMax-M2, so the hosted $0.30 per Mtok input price is a convenience rather than a lock-in.
- is minimax-m2 good for agentic coding
- MiniMax-M2 uses interleaved thinking with explicit <think> tags and was trained for coding-run-fix loops and test-validated repairs, which maps directly onto Atlas's build-and-verify cycle.
- should i upgrade from minimax-m2 to a newer minimax model
- MiniMax-M2.1 through M2.7 hold the same $0.30 and $1.20 pricing with a larger 204,800 token window, so there is no cost penalty for moving off MiniMax-M2.
Try Atlas in your terminal
The terminal-native AI coding agent. Free core, single binary.
Install AtlasRelated guides
Atlas for Symfony in 2026
Atlas is a terminal-native AI coding agent for Symfony in 2026. It reads autowired services and Doctrine mappings, and shows migration SQL before anything runs.
Rename a Symbol Across the Repo with Atlas in 2026
How to rename a symbol across a repo with Atlas in 2026: findReferences gets the true reference set, grep catches strings and docs, and edit refuses ambiguous matches.
Atlas vs GitHub Copilot: Terminal AI Coding Agents in 2026
Comparing Atlas, the terminal-native AI coding agent, with GitHub Copilot's editor extension and chat features for developers in 2026. Explore planning, pricing, and privacy.
Atlas for Astro: Islands, Content Collections, and Zero JS by Default in 2026
Atlas is a terminal-native AI coding agent for Astro in 2026. It reads astro.config.mjs, src/pages, and content collection schemas, drops needless client:load directives, and runs astro check.
Onboard to an Unfamiliar Codebase with Atlas in 2026
How to onboard to an unfamiliar codebase with Atlas in 2026: use codebase_search, glob, read, lsp, task, and todowrite to build a mental model fast.
Plan a Multi-File Change Before Editing with Atlas in 2026
How to plan a multi-file change with Atlas in 2026: the plan agent denies all edit tools, you research with codebase_search and lsp, then plan_exit hands off.
Atlas vs Fine.dev: Terminal AI Coding Agents in 2026
Compare Atlas and Fine.dev for terminal AI coding in 2026. Atlas offers a free core TUI with permission-gated tools, while Fine.dev provides asynchronous cloud agents and prebuilt libraries.
Atlas vs Qwen Code: Terminal AI Coding Agents in 2026
Atlas and Qwen Code are leading terminal AI coding agents in 2026. Compare Atlas's TUI, permission-gated tools, and free core with Qwen Code's 1M token context, Agent Teams, and Alibaba Cloud hosting.