MiniMax open-sources M2.7 with SOTA coding scores

MiniMax M2.7 is a Mixture-of-Experts coding and agent model that just had its weights officially released on Hugging Face. If you have been looking for an open-weight foundation capable of handling real-world software engineering, terminal execution, and multi-agent workflows without getting locked into closed APIs, this one is worth a close look.

First unveiled on March 18, M2.7 is MiniMax’s most capable open-source release to date. Because it uses a MoE design that activates only a fraction of its total parameters during each forward pass, it runs faster and costs less to serve than dense architectures of similar output quality.

What it can do

The model centers on three key capabilities: professional software engineering, office productivity, and native multi-agent collaboration, which MiniMax calls Agent Teams. It is designed to navigate complex tool harnesses, search for external tools dynamically, and orchestrate sub-agents.

On the benchmarks, M2.7 scored 56.22% on SWE-Pro, matching GPT-5.3-Codex. Unlike academic coding tests, SWE-Pro covers messy tasks like log parsing, incident debugging, security auditing, and machine learning pipelines across multiple programming languages. It also hit a state-of-the-art 57.0% on Terminal Bench 2 and scored 39.8% on NL2Repo.

For repository-level generation, it reached 55.6% on VIBE-Pro, closely trailing Opus 4.6 across Web, Android, iOS, and simulation tasks. It also posted 76.5 on SWE Multilingual and 52.7 on Multi SWE Bench.

Beyond basic code completion, MiniMax tested M2.7 on live production alerting. The model correlated metric anomalies with deployment windows, sampled tracing logs, queried production databases to verify root causes, found missing migration files, and drafted non-blocking index fixes to halt live incidents in under three minutes.

Self-improvement and benchmarks

M2.7 is also MiniMax’s first model to work on its own development loop. In an internal self-evolution experiment, the model ran over 100 autonomous cycles of analyzing failure trajectories, updating scaffold code, running evaluations, and committing improvements.

Through this loop, M2.7 independently uncovered optimal sampling hyperparameter configurations, authored new workflow heuristics like automated cross-file bug pattern sweeps, and introduced loop-detection mechanisms to the scaffold. That process drove a 30% jump on internal evals, and MiniMax reports the model now manages 30% to 50% of the workflow inside its internal reinforcement learning team.

On OpenAI’s MLE Bench Lite suite of 22 competitions runnable on a single A30 GPU, M2.7 secured 9 gold, 5 silver, and 1 bronze medals in its top 24-hour run. Its average medal rate of 66.6% ties Gemini-3.1, trailing only Opus-4.6 at 75.7% and GPT-5.4 at 71.2%.

How to get it

The model weights are available right now on Hugging Face under the MiniMaxAI organization.

For teams that prefer managed inference over provisioning MoE clusters, MiniMax is also serving M2.7 through its official developer platform API.

Why it matters

Software teams and DevOps engineers exploring self-hosted agent stacks should give M2.7 a serious spin. Hitting parity with frontier models like GPT-5.3-Codex on realistic engineering tasks has historically been out of reach for open weights.

That said, deploying an MoE model of this caliber still requires serious GPU hardware and a refined execution sandbox. The benchmark numbers look stellar, but turning raw terminal comprehension into safe, automated production fixes will depend heavily on the safeguards developers wrap around it.