MiniMax-M2 sets an open-weights high on a leading index, with caveats

This week the open-weights field has a new reference point. Artificial Analysis, an independent benchmarking service, reported that MiniMax's M2 "achieves a new all-time-high Intelligence Index score for an open weights model" ^[1]. For businesses weighing self-hosted against hosted AI, the details matter more than the headline.

What was released

MiniMax published the weights of M2, a mixture-of-experts model built for coding and agent workflows, on or about October 27, 2025: the model card says "Today, we release and open source MiniMax-M2" without a date, and Artificial Analysis's analysis is dated October 27 ^[1][2]. Artificial Analysis lists 200B total parameters with 10B active per forward pass; MiniMax's own model card says 230B total with 10B active (its file metadata shows about 229B) ^[1][2]. We take the vendor's 230B as the more likely figure, but the two sources disagree, so verify before sizing hardware. In a mixture-of-experts design only a fraction of the model runs for each token, which keeps serving cost closer to that of a small model.

The model card shows the Artificial Analysis Intelligence Index at 61 and 69.4 on SWE-bench Verified, a software-engineering benchmark ^[2]. The card's license field reads "modified-mit" while the page body links to an MIT license, so read the license text itself ^[2].

Price and efficiency

Through MiniMax's own API, M2 costs $0.30 per million input tokens and $1.20 per million output tokens ^[1]. Anthropic's Claude Sonnet 4.5, released a month earlier, is priced at $3 and $15 ^[3]. The model is verbose, however: it used about 120 million tokens to complete Artificial Analysis's index evaluations, tied for the highest of the models it compared ^[1]. Cost per task, not per token, is the number to test.

The SWE-bench figures come from each vendor under its own test setup and are not strictly comparable. The Sonnet 4.5 number is Anthropic's primary score averaged over 10 trials ^[3].

Caveats

Artificial Analysis noted that models such as DeepSeek V3.2 and Qwen3 235B may still outperform M2 at some generalist tasks ^[1]. A strong index score is not a guarantee on your workload, and the license, data-handling and security review of any model you host remain your responsibility.

Practical implication

Open-weights models are now credible candidates for coding and agent tasks where data must stay in-house or per-token cost dominates. Test one against your own tasks, measuring total tokens used and not just the price list.