MiniMax M2: The Open LLM Beating Claude and Gemini!

Summary

Minimax M2, an open 200-billion-parameter mixture-of-experts (MoE) model with only ~10 billion active parameters at inference, is being touted as a frontier alternative that outperforms many proprietary models on key benchmarks. The model ranks fifth on the artificial analysis benchmark, shows strong coding and agent/tool-use performance, and claims faster inference through sparse activations. Minimax can be run locally (reportedly on four GPUs via vLLM) or accessed via a low-cost API, and the makers publish recommended inference settings. The team also offers free access on its platform for trial use.

Original Description

MiniMax just dropped a monster: the MiniMax M2, an open-weight 230B MoE model with 10B active parameters — think Claude Sonnet speed at 8% of the price, and it’s open.

It ranks #5 globally on the AA benchmark and still #1 among open models, outperforming Claude Sonnet 4 and Gemini 2.5 Pro in both agentic and coding tasks.

If you’ve been hunting for frontier-level performance without frontier-level costs, this might be your new go-to:

⚙️ 200K context

⚡ ~2x faster inference

💡 optimized for agent loops and reliability

Oh, and it runs locally on just 4×H100s. This is what open AI done right looks like.

I’m Louis-François, PhD dropout, now CTO & co-founder at Towards AI. Follow me for tomorrow’s no-BS AI roundup 🚀

#minimax #aiagents #opensourceai #short