MiniMax-M2.1

What it is

MiniMax-M2.1 is an open-weight large language model from MiniMax, published at the end of 2025 as the successor to the company's M2 model. The public checkpoint is listed at 229B parameters ^[1]. The weights were posted to Hugging Face on 20 December 2025, and a Hugging Face blog post about the release carries a date of 5 January ^[2][1].

The weights are released under a modified MIT license. The same model is served through an API on the MiniMax Open Platform, and a coding product called MiniMax Code is publicly available ^[1]. So the release offers two ways in: download the weights and run them yourself, or call the hosted version.

Background and positioning

MiniMax described M2.1 as a significant leap over M2 on software-engineering leaderboards, with particular emphasis on multilingual coding ^[1]. That framing shaped which benchmarks the company reported: alongside the general software-engineering tests it also published multilingual coding suites and mobile and web development suites, where the gap to the previous generation is widest.

Benchmark results

On the multilingual software-engineering benchmarks, SWE-bench Multilingual rose from 56.5 for M2 to 72.5, ahead of Claude Sonnet 4.5 at 68.0, while Multi-SWE-bench rose from 36.2 to 49.4 ^[1]. These tests ask a model to resolve real issues in code repositories across many programming languages; each task is scored by whether the resulting patch passes the project's tests.

On SWE-bench Verified, a curated set of software-engineering tasks, the score went from 69.4 for M2 to 74.0. Terminal-bench 2.0, which measures work carried out through a command-line environment, rose from 30.0 to 47.9 ^[1].

The VIBE suite, which covers web, Android and iOS application development, showed the largest movement: the average went from 67.5 to 88.6, with VIBE-Web at 91.5 and VIBE-Android at 89.7 ^[1]. VIBE-iOS rose from 39.5 to 88.0 ^[1].

Two further results point at the same direction. SWT-bench, which concerns writing tests, went from 32.8 to 69.3 ^[1]. On SWE-bench Verified run through the mini-swe-agent harness — a different agent setup around the same model — the score was 67.0, up from 61.0 for M2 ^[1].

BrowseComp, a benchmark for finding hard-to-locate information on the web, went from 44.0 to 47.4, and reached 62.0 with context management, against 56.9 for M2 ^[1].

Comparison with the Claude models

MiniMax said M2.1 outperforms Claude Sonnet 4.5 and closely approaches Claude Opus 4.5 in multilingual scenarios ^[1]. In the company's own comparison table the picture is more mixed. On SWE-bench Verified, M2.1 scored 74.0, against 77.2 for Claude Sonnet 4.5 and 80.9 for Claude Opus 4.5. On Multi-SWE-bench it scored 49.4, against 50.0 for Claude Opus 4.5 ^[1].

In other words, the claim of an advantage rests on the multilingual coding benchmarks, where M2.1 is placed ahead of Sonnet 4.5, while on the single-language Verified benchmark the same table places it behind both Claude models ^[1].

Running the model

The published recommended inference settings are temperature 1.0, top_p 0.95 and top_k 40 ^[1]. For serving, the listed frameworks are SGLang, vLLM, Transformers, KTransformers and MLX-LM ^[1] — a spread that covers datacenter serving, ordinary Python inference, CPU-oriented and Apple-silicon setups.

Where it sits

M2.1 continues MiniMax's line of open-weight models and pairs the weights with a hosted API and a coding product ^[1]. Its reported gains are concentrated in agentic, tool-using work — resolving repository issues, operating a terminal, browsing the web, building mobile and web apps — and in the multilingual versions of those tasks ^[1]. That is where the model is meant to be judged.