ModelsMiniMax M2 Guide: The 10B-Active Model That Started the Line
MiniMax M2 put a 230B MoE with about 10B active behind a cheap API in October 2025. What it is, the thinking-history trap, and where it still earns a place.
MiniMax M2 is now a legacy model on MiniMax's own platform, and it has been overtaken several times within its own family. It is still worth understanding, for two reasons. It set the pattern every later M2-series model follows, and it has one integration requirement that quietly costs people quality when they get it wrong.
This guide covers what M2 is, the mistake most integrations make with it, and the narrow set of cases where it is still a sensible choice.
What M2 is
MiniMax announced M2 on 27 October 2025 as a model for agents and code, and published the full weights on Hugging Face at the same time. The model card describes a mixture-of-experts model with about 230B total parameters and about 10B active per token, under a licence MiniMax labels modified MIT.
The launch pitch was price and speed rather than leadership. MiniMax listed the API at $0.30 per million input tokens and $1.20 per million output, which it framed as 8% of the price of Claude Sonnet 4.5, and quoted online inference at around 100 tokens per second. It was candid that coding was slightly behind the best closed models while tool use and deep search were close to them.
The vendor-reported scores on the model card fit that description:
- SWE-bench Verified — 69.4
- SWE-bench Multilingual — 56.5
- Multi-SWE-Bench — 36.2
- Terminal-Bench — 46.3
- BrowseComp — 44.0
- τ²-Bench — 77.2
MiniMax's current documentation lists a 200K context window and a maximum output of 128K tokens, with the note that the output budget includes chain-of-thought. OpenRouter listed it at $0.255 input and $1.02 output per million tokens at the time of writing.
Small active count, large memory footprint
About 10B active parameters is very small for a model in this class. It is why M2 was fast and cheap to serve: compute per token scales with the active count, not the total.
Memory does not. All 230B weights have to be resident to serve the model, because any token can be routed to any expert. That is the gap people miss when they read "10B active" and imagine running it on one consumer card. Self-hosting M2 at useful precision is a multi-GPU job, or a heavily quantised one with the quality loss that implies. Active vs total parameters works through the arithmetic, and best model for self-hosting covers what fits on which hardware.
Do not strip the thinking
This is the practical point that matters most. M2 is an interleaved thinking model: it reasons inside <think>...</think> tags between tool calls, and the model card is explicit that the thinking content must be kept in the conversation history when you send earlier turns back. Removing it, MiniMax warns, degrades performance.
Plenty of client code strips reasoning blocks by default, to save tokens or to keep transcripts tidy. With most models that is harmless. With M2 it means every new turn starts without the reasoning that produced the previous tool call, and agent loops get noticeably worse at following their own plan.
Check three places:
- Your agent framework. Some drop reasoning content from assistant messages before re-sending history.
- Any proxy or gateway in between. Some normalise responses and discard non-standard fields.
- Your own context compaction. A summariser that throws away thinking blocks has the same effect, just later in the session.
The cost is real: preserved thinking means more input tokens per turn. That is the trade MiniMax designed for, and it is cheaper than a loop that loses the thread. MiniMax also recommends temperature 1.0 and top_p 0.95; very low temperatures, a common default for coding, are not what it was tuned for.
What M2.5 changed
The family moved quickly. M2.1 followed M2, and M2.5 arrived in February 2026 with MiniMax reporting 80.2% on SWE-bench Verified against M2's 69.4, plus large gains on Multi-SWE-Bench and BrowseComp. Those are vendor numbers and the scaffolds differ, but the direction is not in doubt.
For new coding or agent work, M2 is rarely the right starting point. The MiniMax M2.5 guide covers the successor, including the fact that MiniMax is retiring M2.5 on its own API in November 2026, and the MiniMax M2.1 guide covers the step in between.
Where M2 still fits
There are a few legitimate reasons to keep running it:
- A frozen, validated pipeline. If you have an eval suite that passes on M2 and the work has not changed, a swap buys risk without a clear return.
- A fine-tune built on it. Open weights mean you may have adapted it. Moving means redoing that work.
- A self-hosted deployment sized for it. If your hardware is already provisioned for M2's footprint, a later model with a different shape is an infrastructure change, not a config change.
Outside those cases, treat M2 as a reference point. Run your own tasks against it and a newer model on the same harness, compare cost per completed task, and move once the newer model wins on your work rather than on a chart.
Common questions
How many parameters does MiniMax M2 have?
The model card describes about 230B total parameters in a mixture-of-experts design, with about 10B active per token. Compute per token follows the active count, but all of the weights must be in memory to serve it.
Why does MiniMax M2 perform worse in my agent framework?
The most common cause is stripped reasoning. M2 is an interleaved thinking model and MiniMax says the think-tag content must stay in the conversation history. Frameworks or proxies that remove it degrade multi-turn performance.
Is MiniMax M2 open source?
The full weights are on Hugging Face under a licence MiniMax labels modified MIT. That is open weights rather than open source in the strict sense, and the modifications are worth reading before commercial use.
