Four days after launching M2.7 as an API-only product, Chinese AI lab MiniMax has confirmed the model will get an open-weights release - meaning anyone can download and run it locally.
That matters because M2.7 is not a mid-tier model getting a publicity bump. It scored 56.22% on SWE-Pro (a software engineering benchmark), putting it neck-and-neck with GPT-5.3-Codex and just a point behind Claude Opus 4.6. On SWE-bench Verified, it hit 78%, significantly outperforming Opus's 55%. The Artificial Analysis Intelligence Index ranked it #1 out of 136 models in its class.
A Model That Helped Train Itself
M2.7's headline feature is what MiniMax calls "self-evolution." During training, the model ran over 100 rounds of its own optimization - analyzing where it failed, modifying its own training scaffolding, running evaluations, and deciding whether to keep or revert changes. MiniMax says the model handled 30-50% of its own reinforcement learning workflow without human intervention, producing a 30% improvement on internal benchmarks.
The architecture follows the same Mixture-of-Experts (MoE) design as earlier M2 models: roughly 229 billion total parameters, but only about 10 billion active during any given computation. That keeps inference costs low. API pricing sits at $0.30 per million input tokens and $1.20 per million output tokens - roughly 50x cheaper than Claude Opus.
What Open Weights Means for Local Users
MiniMax has a track record here. M2, M2.1, and M2.5 all shipped with open weights under MIT or Modified-MIT licenses. The M2.5 model, released in February 2026, already runs locally and outperforms Llama 3.1 405B on coding benchmarks while using a fraction of the compute (only 10B active parameters vs. 405B).
If M2.7 follows the same licensing pattern, it would be the most capable open-weights model available by a significant margin. For people running local AI setups, that is a concrete upgrade - frontier-level coding assistance, document processing, and multi-agent orchestration without API costs or data leaving your machine.
There are tradeoffs. M2.7 generates output at 45.5 tokens per second, which is slow compared to peers. It is also extremely verbose - during benchmark evaluations, it produced roughly 87 million output tokens versus a 20 million median. Time to first token averages 2.21 seconds. These quirks are manageable on an API but could be more noticeable running on local hardware.
MiniMax's Trajectory
MiniMax, founded in 2021 as a SenseTime spinoff, went public in Hong Kong in January 2026, doubling on its debut to a $12.8 billion market cap. The company has raised over $850 million from Alibaba, Tencent, and MiHoYo. Their consumer products include the Talkie AI character app and Hailuo video generation tool.
No release date for the open-weights version has been announced. Given that M2.5 went from launch to open release in roughly the same timeframe, an open M2.7 could arrive within weeks.