MiniMax releases MiniMax-M2.1 for multilingual coding

Background

MiniMax measured M2.1 against its predecessor M2, the model it had used as the baseline for coding progress. The release is aimed at software-engineering work that crosses programming languages, and at tasks where a model has to call tools rather than only write code.^[1]

Release

Weights for M2.1 were posted to Hugging Face on 20 December 2025; the Hugging Face blog post about the model is dated 5 January.^[2][1] MiniMax's own release post is titled "M2.1: Multilingual and Multi-Task Coding with Strong Generalization".^[1] The weights are open under a modified MIT license. The model is also offered through the API on the MiniMax Open Platform, and MiniMax Code, the company's coding product, was made publicly available at the same time.^[1]

Benchmark results

MiniMax described the release as a significant leap over M2 on software-engineering leaderboards, with the largest gains in multilingual coding.^[1] On SWE-bench Verified, which asks a model to resolve real issues taken from open-source repositories, M2.1 scored 74.0 against 69.4 for M2.^[1] On SWE-bench Multilingual, the same kind of work spread across many programming languages, the score rose from 56.5 to 72.5, ahead of Claude Sonnet 4.5 at 68.0.^[1] Multi-SWE-bench went from 36.2 to 49.4, and Terminal-bench 2.0, which tests work inside a command-line environment, from 30.0 to 47.9.^[1]

Two benchmarks closer to agent work moved further. BrowseComp, which measures finding information on the web, scored 47.4 against 44.0 for M2.^[1] Toolathlon, which tests tool use, rose from 16.7 to 43.5.^[1] MiniMax said the model outperforms Claude Sonnet 4.5 and closely approaches Claude Opus 4.5 in multilingual scenarios.^[1]

Running the weights

The Hugging Face model card recommends SGLang, vLLM, Transformers, KTransformers and MLX-LM for serving the model.^[1]

Why it matters

The distance between SWE-bench Verified and SWE-bench Multilingual is one way to read how well a model handles code outside the most common languages and ecosystems. M2 sat at 69.4 and 56.5 on the two; M2.1 sits at 74.0 and 72.5, so the gap narrows sharply, which matches the multilingual emphasis MiniMax put on the release.^[1] The jump on Toolathlon, from 16.7 to 43.5, concerns a different skill: calling tools correctly as part of a task, which is what a coding agent spends much of its time doing.^[1]

Open weights under a modified MIT license mean researchers and companies can download and run the model themselves rather than reaching it only through an API, and the model card lists several established inference frameworks for that purpose.^[1] For users, M2.1 is reachable in three ways at once: the open weights, the platform API, and the MiniMax Code product.^[1]