Direct answer
MiniMax M2 may be worth evaluating for coding and tool-using workflows, based on historical descriptions and third-party evaluation. But a current direct MiniMax API route for M2 could not be verified from the reviewed documentation. MiniMax’s API overview names M3.1-Flash and M3; that alone does not establish that its list is exhaustive or that M2 is retired. Don’t commit to a new M2 integration without confirming access and terms with the provider. (MiniMax API overview)
An October 2025 Ollama post showed an M2 cloud API example using Ollama’s endpoint and credentials. That documents a historical third-party route, not MiniMax’s direct API. Ollama’s library separately marks its own M2 offering retired as of June 16, 2026; that status applies to Ollama’s offering, not necessarily to M2 everywhere. (Ollama’s M2 post; Ollama’s model page)
What MiniMax M2 was designed for
MiniMax introduced M2 in 2025 as an open model for coding and agent workflows. MiniMax’s model repository reports 230 billion total parameters and 10 billion active parameters. Those figures differ from Artificial Analysis’s October 27, 2025 report of 200 billion total parameters and 10 billion active parameters. The available information does not explain the difference, so treat the total-parameter count as disputed rather than assuming the reports are interchangeable. (MiniMax’s M2 repository; Artificial Analysis)
Artificial Analysis identified tool use and instruction following as strengths in its evaluation, while cautioning that M2 might lag some open models on certain generalist tasks. Ollama also described M2 as suited to coding and agentic workflows, including multi-file edits and long tool chains. These are historical publisher claims; the available descriptions do not provide enough visible scores and methodology to establish how M2 would perform on your specific workload. (Artificial Analysis; Ollama)
Treat coding, tool use, and agent flows as candidates to test, not guaranteed strengths. If your workload requires multimodal input, note that Artificial Analysis characterized M2 as text-only in its October 2025 analysis; verify current capabilities before relying on that dated assessment. (Artificial Analysis)
Is there a production API for M2?
The reviewed MiniMax API overview names M3.1-Flash and M3, but the information available here does not establish whether that list is exhaustive or confirm a current direct API route for M2. MiniMax’s invocation documentation describes an Anthropic-compatible route for a currently documented model; consult that documentation for the current model ID, authentication, and request details. Do not substitute an M3.1-Flash example for an M2 integration. (MiniMax API overview; MiniMax invocation documentation)
The older Ollama example used https://ollama.com/api/chat, an Ollama API key, and the model name minimax-m2. It was an Ollama-hosted example—not a MiniMax endpoint or evidence of current availability. Ollama’s model page now marks its M2 offering retired. (Ollama’s M2 post; Ollama’s model page)
Cost: don’t treat old M2 prices as a current quote
Artificial Analysis reported historical M2 API rates of $0.30 per 1 million input tokens and $1.20 per 1 million output tokens in its October 27, 2025 article. These are dated third-party figures, not a current M2 price. MiniMax’s pricing page lists M3.1-Flash at $0.15 per 1 million input tokens and $0.60 per 1 million output tokens; check the page for applicable rates and terms before budgeting. (Artificial Analysis; MiniMax pricing)
For illustration, at those listed M3.1-Flash rates, 1 million uncached input tokens plus 200,000 output tokens would cost $0.27 before any other applicable charges: $0.15 for input and $0.12 for output. This is arithmetic using the listed rates, not a measured workload result. (MiniMax pricing)
Artificial Analysis also reported that M2 used 120 million tokens in its own Intelligence Index evaluations. That is an evaluation-specific figure, not a forecast for your prompts. Compare total cost per successful task using your own token use, task outcomes, and current provider terms—not per-token prices alone. (Artificial Analysis)
How to decide whether an alternative fits
Run a small, repeatable comparison before adopting a model:
- Choose representative tasks. Use real coding tickets, multi-file edits, tool calls, or research flows from your workload.
- Keep conditions consistent. Use the same inputs, prompts, tools, permissions, and success criteria across candidates.
- Score the outcome. Track task completion, correctness, tool-call failures, and human repair required.
- Measure cost and speed. Record input and output tokens and end-to-end latency under expected conditions. Don’t substitute benchmark token counts for your own.
- Confirm the route. Before implementation, verify the provider’s current model ID, endpoint, authentication, limits, and billing terms.
Bottom line: M2’s historical coding and agentic positioning may make it a candidate for workload testing, but a current direct MiniMax API route for M2 could not be verified here. The reviewed MiniMax documentation names M3.1-Flash and M3; confirm availability directly before integrating M2, and choose based on representative task results rather than old prices or rankings.
Sources
API Overview — MiniMax API Docs
Model Invocation — MiniMax API Docs
Pricing Overview — MiniMax API Docs
MiniMax-M2 — MiniMax-AI GitHub