MiniMax-M2.7

What it is

MiniMax-M2.7 is a large language model released by MiniMax in March 2026.^[2] Its paper is arXiv 2605.26494, and the published checkpoint is listed at 229 billion parameters; the Hugging Face model card gives the same 229B total and a license field that reads "other".^[1] The size figure matters mainly as context: it tells readers what class of model this is, a large model meant to be run as a checkpoint rather than a small model embedded in a product.

The model sits in a line of work aimed less at single-turn question answering than at sustained task execution, where a model is expected to call tools, run code, and keep working across many steps.

Where it came from

MiniMax described the model as "our first model deeply participating in its own evolution".^[1] The company supported that description with one internal result: a version of the model was given a programming scaffold and optimized it autonomously over more than 100 rounds, reporting a 30% performance improvement.^[1]

A scaffold, in this context, is the surrounding harness that a model works inside, the code that decides how it edits files, runs tests and retries after failures. Having a model rewrite that harness is a way of letting it improve the conditions of its own work rather than only its answers. The claim here is specific and bounded: one internal version, one scaffold, more than 100 rounds, a 30% gain as reported by the company.^[1]

What it can do

The model supports Agent Teams, complex Skills and dynamic tool search, and MiniMax also points to stronger character consistency for entertainment uses.^[1] Agent Teams refers to splitting a job among several cooperating model instances; Skills refers to packaged procedures a model can load and follow; dynamic tool search means the model looks up which tools are available instead of being handed a fixed list up front. The character-consistency point is about keeping a persona stable across a long interaction, which matters for interactive entertainment products.

Benchmark results

MiniMax reported SWE-Pro at 56.22%, Terminal Bench 2 at 57.0%, SWE Multilingual at 76.5, GPQA Diamond at 89.8 and AIME 2026 at 94.2.^[1] On MLE Bench Lite, a set of 22 machine-learning competitions, the model reached a 66.6% medal rate, which MiniMax said was second only to Opus-4.6 and GPT-5.4.^[1]

The Hugging Face model card carries a longer list. It gives Humanity's Last Exam at 28 and MMLU-Pro at 81.8; SkillsBench V1.1 at 34.9, Multi SWE Bench at 52.7 and VIBE-Pro at 55.6%; Toolathon at 46.3% accuracy, NL2Repo at 39.8 and MM Claw at 62.7%; and GDPval-AA at an ELO of 1495.^[1]

Read together, the numbers describe a model that is strong on competition-style mathematics and science questions and on software-engineering tasks, and noticeably weaker on the agentic and tool-use benchmarks, where scores in the thirties and forties appear alongside the higher figures. The gap between the two groups is the most informative thing in the table: it shows where the difficulty has moved once a model is judged on multi-step work rather than on single answers.

Where it sits in the line

The model was followed by MiniMax-M3 in June 2026.^[2] That short interval places M2.7 as one step in a rapid sequence of releases rather than as an endpoint, and the self-improvement framing MiniMax used for it is best read in that light, as a description of a direction the company was pursuing at that point in the sequence.