MiniMax-M2 is an open-weight mixture-of-experts (MoE) text generation model from MiniMax that’s explicitly framed around “agentic” and coding-heavy workloads. The model card describes it as 230B total parameters with 10B active parameters, aiming to keep per-token compute closer to a much smaller model while still benefitting from a large expert pool.
If you’re evaluating it, the most interesting angle is less “can it chat” and more “can it reliably do end-to-end developer loops”: multi-file edits, reasoning across long contexts, and tool-driven workflows. The authors’ README emphasizes performance on coding and agent benchmarks, and notes evaluations run with very long context windows (reported as 128k in their SWE-bench Verified setup). Practically, you’ll want to test it on a small repo task that has an objective pass/fail signal (tests or a linter) and see whether the model can converge without excessive retries.
Quick stats from the listing feed: pipeline: text-generation · 1416 likes · 131385 downloads.
Source listing: https://huggingface.co/models?sort=modified