MiniMax open-sources M2.7 agent model with 56.22 on SWE-Pro
TL;DR
MiniMax open-sourced M2.7 on April 12, 2026, scoring 56.22 on SWE-Pro and 57.0 on Terminal-Bench 2.0. License is more restrictive than the MIT terms used for M2 and M2.5.
What changed
MiniMax released M2.7 publicly on April 12, 2026 (originally announced March 18). Headline benchmarks: SWE-Pro 56.22 and Terminal-Bench 2.0 57.0. The release introduces a self-evolving architecture in which the model reviews its own outputs and feeds revisions back into the training signal. Distribution is via HuggingFace weights, NVIDIA platform integration, and the MiniMax direct API.
Why it matters
M2 and M2.5 shipped under MIT with zero commercial restrictions. M2.7 breaks that pattern with a more restrictive license, mirroring Alibaba's Qwen 3.6-Max pivot to closed and tightening the open-frontier field to Moonshot Kimi, Z.AI GLM-5.1, DeepSeek V4, and Google Gemma 4. The M-series remains specifically built for agentic harnesses (multi-step tool use, long-running autonomous workflows) rather than general chat, so the SWE-Pro and Terminal-Bench scores are the ones to weigh against your harness latency and throughput targets.
What to watch for
Read the new license terms before deploying: if restrictions are limited to fine-tune-and-redistribute scenarios, deployment-only API integrations remain unaffected. Track M2.8 benchmarks in two to three months to verify whether the self-evolving claim produces compounding gains or is one-time uplift. Watch whether DeepSeek and Moonshot follow MiniMax toward more restrictive flagship licensing.
Who this matters for
- Developers: 56.22 on SWE-Pro and 57.0 on Terminal-Bench 2.0 make M2.7 competitive with closed alternatives at lower $/M tokens. Read the new license carefully before committing production agent workloads.
Harsh’s take
The license change is the story. For 18 months, MiniMax was the predictable open-weight Chinese lab: MIT on everything, no strings, drop in and use. M2.7 breaking that pattern signals the economics of pure open-source distribution are getting harder even for labs that built their reputation on it. DeepSeek and Moonshot remain fully open for now; Z.AI is MIT on GLM-5.1. The trend across Chinese frontier labs is clear: monetise the flagship, keep older generations open as distribution fuel.
The self-evolving framing is the architectural claim worth watching. Most self-improving labels applied to models so far have been marketing for standard RLHF. If M2.7 actually demonstrates compounding capability gains through iterative self-review rather than just scoring better on one benchmark, that is a different class of capability. Real test: M2.8 benchmarks relative to M2.7 in two to three months.
by Harsh Desai
About MiniMax Agent
View the full MiniMax Agent page →All MiniMax Agent updatesMore AI news
- Daily RoundupGoogle magenta-realtime-2 and gemma-4-12B trend on Hugging Face plus new Replicate characters and Seedance video on Fal
Google placed two new models at the top of Hugging Face trends while small character generators appeared on Replicate and ByteDance dropped a faster reference-to-video option on Fal, giving builders quicker access to audio, multimodal, and video tools.
- Daily RoundupVercel Sandbox drives, Gemma 4 QAT, and NVIDIA models trending on Hugging Face (agent tools today)
Vendors added persistent storage and quantized models while new text-to-speech and image models appeared on Hugging Face and Fal, with fresh agent tools listed on Product Hunt.
- Weekly DigestCursor enterprise orgs, Claude Code Bedrock auto mode, Codex Sites preview for quick deploys
Cursor added team management and design tools while Claude Code and Codex rolled out cloud integrations, safety checks, and deployment features across CLI and apps from 29 May to 5 June 2026.