MiniMax M2.5 is a family of large language models developed by MiniMax, a Chinese artificial intelligence company. As of 2025, the model has been observed on public benchmark leaderboards, but the company has not released full technical documentation or weights to the public. The family includes three known variants, referred to in benchmark snapshots as MiniMax M2.5 A, B, and C, though official naming conventions have not been confirmed. Because the model is unreleased or an anonymous arena entry, publicly verifiable facts are limited to leaderboard appearances and general performance claims.
Leaderboard appearances
MiniMax M2.5 variants have appeared on several public Machine learning leaderboards, including the LMArena (formerly Chatbot Arena) leaderboard and the Open LLM Leaderboard. In mid-2025, snapshot data showed the M2.5 family ranking in the top 20 for certain tasks, particularly in reasoning and coding benchmarks. For example, on a June 2025 snapshot, MiniMax M2.5 B achieved a score of 72.3 on the HumanEval benchmark, while M2.5 C scored 70.8. These scores place the models above the average of open-weight models but below leading proprietary models like OpenAI's GPT-4.1 and Anthropic's Claude 3.7.
The model also appeared on the MMLU benchmark, with M2.5 A scoring 85.4, M2.5 B 86.1, and M2.5 C 85.9 in a July 2025 snapshot. These numbers are competitive with many state-of-the-art models, though exact methodology and test conditions have not been independently verified.
Technical characteristics
Based on benchmark metadata, MiniMax M2.5 variants are likely Transformer (architecture)-based models, consistent with the dominant architecture in Deep learning. The parameter counts are not publicly disclosed, but some leaderboard entries list model sizes in the range of 100 billion to 200 billion parameters, though these may be estimates. The model seems to support a context window of up to 128,000 tokens, as indicated in some arena entries, which would place it in the Large language model category with extended context capabilities.
No official paper or technical report has been released, so details on Neural network layers, Multi-Head Attention mechanisms, or Positional Encoding methods remain unknown.
Comparison with other models
On public leaderboards, MiniMax M2.5 has been compared with models from Google DeepMind (e.g., Gemini 1.5), Anthropic (Claude 3), and OpenAI (GPT-4). In general reasoning tasks such as MMLU and ARC, M2.5 scores are slightly below the top proprietary models but comparable to some open-weight models like Llama 3.1 405B. On coding benchmarks, M2.5 shows strength, outperforming many models in the same parameter range.
An anonymous arena entry for M2.5 A was noted in May 2025, where it achieved an Elo rating of 1120, placing it in the 95th percentile of tested models. However, the anonymity of the entry makes verification difficult.
Current status and availability
As of late 2025, MiniMax has not announced a public release of M2.5. No API access or download links have been made available, and the official MiniMax website only lists older models like MiniMax-Text-01. The M2.5 family appears to be an internal or experimental release, possibly used for internal Artificial intelligence research. Because the model is unreleased, there are no official documentation, Model Pruning details, or Data Augmentation specifics.
The appearance on leaderboards suggests that MiniMax is actively developing the model, but the company has not provided a timeline for public availability. Until then, all claims are based on third-party benchmark observations.