Short answer
MiniMax M2 is an open-weight model aimed at coding and agentic workflows, including multi-step tool use. MiniMax describes it as a mixture-of-experts model with 230 billion total parameters and 10 billion active parameters. For production, distinguish running M2 yourself from using a managed MiniMax API: MiniMax publishes a vLLM deployment guide for M2, but the current API documentation reviewed here lists M3.1-Flash and M3—not original M2—as supported language models. M2 may merit evaluation if you can operate the required GPU infrastructure; don’t assume a managed-API integration is available. (M2 repository; API invocation docs)
What is MiniMax M2?
This guide covers original MiniMax-M2, not M2.1, M2.5, M2.5-Lightning, or M2.7. MiniMax announced M2 on October 27, 2025, as a model for agents and coding. Its official repository describes an MoE architecture with 230 billion total parameters and 10 billion active parameters. Those are model specifications, not a promise that serving it will be inexpensive: the official deployment guide estimates about 220 GB of memory for the weights. (MiniMax announcement; M2 repository; vLLM deployment guide)
MiniMax positions M2 for software development and agents that use tools such as a shell, browser, retrieval system, or code runner. The repository describes coding tasks including multi-file changes, code-run-fix loops, and test-validated repairs. Treat these as the publisher’s intended-use and capability claims; your results will depend on the task, serving setup, harness, and tools. (M2 repository)
What do the published benchmarks indicate?
MiniMax’s repository reports these M2 scores: 69.4 on SWE-bench Verified, 36.2 on Multi-SWE-Bench, 56.5 on SWE-bench Multilingual, and 46.3 on Terminal-Bench. The README presents these in a coding and agentic benchmark table and describes the suite as testing end-to-end tasks such as editing repositories, executing commands, browsing, and delivering solutions. These are MiniMax-published results, not independent verification or a guarantee of performance on your codebase. The table excerpt does not specify a harness or averaging procedure, so treat the scores as a reason to run a representative pilot—not as a substitute for measuring your own pass rate, tool-call reliability, latency, and cost. (M2 README)
Is M2 a good fit for your workload?
Consider testing it if your workload involves coding in a terminal or IDE, multi-step agent tasks, or structured tool use—and if you can support self-hosting or have confirmed another access path. The official tool-calling guide documents an OpenAI-compatible example in which the model selects a weather function and returns structured arguments. That establishes a documented tool-calling workflow for a self-hosted setup, but not that every agent framework or tool schema will work without adaptation. (M2 tool-calling guide)
Be cautious if you need a straightforward, currently documented MiniMax-hosted API, predictable per-token pricing for M2, or managed-service latency guarantees. The current API invocation documentation reviewed here names M3.1-Flash and M3 as supported models, not original M2. MiniMax’s M2 launch API promotion was time-limited and ended November 7, 2025; that promotion is not evidence of current paid access. The documentation reviewed here therefore does not establish a currently supported managed-API path for original M2. (API invocation docs; model release notes)
A practical evaluation before production
Use a small, representative test set rather than relying on a headline score. For example, assemble 30–50 tasks from your own codebase or agent workflow, including successful cases, ambiguous instructions, and tool failures. Run the same prompts and tool setup on M2 and the current alternative you would otherwise choose. Track:
- Task quality: tests passed, correct tool choice, and whether the final result meets your acceptance criteria.
- Reliability: successful completion rate, malformed tool calls, retries, and recovery from tool errors.
- Speed: end-to-end median and 95th-percentile task time under your intended concurrency.
- Operating cost: GPU time, memory use, infrastructure, and engineering effort—not just active parameter count.
These are suggested evaluation measures, not reported M2 test results. Choose M2 only if it meets your workload’s quality and reliability thresholds and its measured operating cost and latency are acceptable.
Can you use MiniMax M2 in production?
Self-hosting: a documented route, with substantial hardware needs
MiniMax publishes a vLLM deployment guide for MiniMaxAI/MiniMax-M2. It specifies Linux, Python 3.9–3.12, GPU compute capability 7.0 or higher, and 220 GB of memory for weights. The guide gives recommended configurations of four 96-GB GPUs and eight 144-GB GPUs; it associates these with supported context lengths up to 400,000 and 3 million tokens, respectively. It also estimates 240 GB of memory per 1 million context tokens and says actual requirements should be adjusted for the use case. The guide does not clarify in the retrieved text whether those context figures are per-request model limits or serving-capacity figures, so confirm that distinction with the deployment documentation and your serving setup before sizing hardware. These configurations indicate a substantial infrastructure requirement, not a lightweight deployment. (vLLM deployment guide)
The same guide documents vLLM launch commands and an OpenAI-compatible local chat-completions test. After following the guide’s installation and serving instructions, run this single-command smoke test against your local server:
curl http://localhost:8000/v1/chat/completions -H "Content-Type: application/json" -d '{"model":"MiniMaxAI/MiniMax-M2","messages":[{"role":"user","content":[{"type":"text","text":"Reply with exactly: M2 ready"}]}]}'
This calls your local vLLM server, not a MiniMax-hosted endpoint. The guide’s example uses http://localhost:8000/v1/chat/completions and the model name MiniMaxAI/MiniMax-M2; follow its current launch instructions for the required serving options. (vLLM deployment guide)
MiniMax’s M2 repository identifies its license as MIT. Review the license and your organization’s security, data-handling, and deployment requirements before using the weights commercially. (M2 repository)
Managed API: verify access before building around it
MiniMax’s current API documentation reviewed here shows how to call currently supported language models and names M3.1-Flash and M3. It does not establish a current M2 model ID, hosted endpoint example, or price. That does not prove no M2 access exists through any channel, but it means the reviewed official documentation is insufficient to recommend a new production integration against a managed M2 API. Confirm the exact model identifier, account eligibility, endpoint, price, limits, and retirement or migration terms with MiniMax before committing to that route. (API invocation docs)
What does MiniMax M2 cost?
The sources reviewed do not establish a current managed-API price for original M2. The M2 launch announcement advertised API calls free for a limited time, ending November 7, 2025; it does not give a current M2 input/output price. Current pricing documentation lists rates for other models, but those rates must not be applied to M2. For self-hosting, compare measured GPU-hours and infrastructure costs against your alternatives; the published weight-memory requirement alone makes clear that hardware is a major part of the decision. (model release notes; pricing overview)
Keep version-specific claims separate. MiniMax’s later M2.5 announcement reports capabilities, benchmarks, speeds, and prices for M2.5 and M2.5-Lightning. Those are not original-M2 specifications or prices. (MiniMax M2.5 announcement)
Bottom line
MiniMax M2 is a coding- and agent-oriented open-weight model with official self-hosting and tool-calling documentation. MiniMax’s published benchmark table makes it a reasonable candidate for a controlled coding or agent pilot, but the results do not establish that it will outperform alternatives on your workload. The decision hinges on deployment: evaluate M2 if you can provide substantial GPU capacity and operate the serving stack; if you need a managed MiniMax API, choose from models currently listed in the API docs or confirm M2 access directly before building against it.
Sources
MiniMax M2 & Agent: Ingenious in Simplicity
MiniMax M2 Model vLLM Deployment Guide
Model Invocation - MiniMax API Docs