MiniMax's new AI model helped build itself

Every engineer jokes about AI taking their job, but MiniMax just turned that concept inward: its new model is already taking over the jobs of the AI researchers who built it.

The Chinese AI lab officially introduced M2.7, a proprietary reasoning model designed to actively participate in its own evolution. Inside MiniMax's reinforcement learning division, M2.7 is already handling between 30% and 50% of the daily experimentation workflow autonomously.

Debugging its own training runs

Instead of just writing boilerplate code, M2.7 operates inside an internal research agent harness across multiple departments.

When a researcher outlines an experimental hypothesis, M2.7 handles literature review, builds data pipelines, and spins up the run. Once the experiment is live, the model monitors logs, runs profiling, pinpoints failures, writes patches, and triggers smoke tests. Human researchers now only step in for high-level directional debates.

In one self-evolution trial, MiniMax let M2.7 autonomously optimize an internal scaffolding environment for code generation. The model ran an iterative loop of analyzing failure trajectories, modifying scaffold code, running evaluations, and keeping or reverting changes for over 100 consecutive rounds. That fully autonomous run yielded a 30% performance boost on internal benchmarks. In live production alerts, it has reportedly dropped incident recovery times to under three minutes.

Punching above its price tag

Across external benchmarks, M2.7 posts numbers right next to top-tier Western models.

On OpenAI's MLE Bench Lite, which simulates 22 machine learning competitions on a single A30 GPU, M2.7 notched an average medal rate of 66.6% across three runs. That ties Google's Gemini 3.1 and sits just behind GPT-5.4 at 71.2% and Claude Opus 4.6 at 75.7%. In its best single run, M2.7 took home nine golds, five silvers, and one bronze.

It scored 56.22% on SWE-Pro, tying GPT-5.3-Codex, and hit 55.6% on VIBE-Pro. It also registered a 34% hallucination rate, beating Claude Sonnet 4.6 and Gemini 3.1 Pro. The pricing is aggressive: $0.30 per million input tokens via the MiniMax API and OpenRouter.

Why it matters

We think the real story here is not another benchmark chart, but how quickly the iterative loop of AI development is closing on itself.

For developers and engineering leads, M2.7 is an immediate, low-cost tool worth testing. At $0.30 per million input tokens, using an agent that can reason through system observability, analyze live database traces, and fix code without hallucinating heavily makes it a very attractive alternative for SRE and backend pipelines.

For the AI industry, this changes the scaling equation. If a model can independently manage nearly half of its own reinforcement learning workflow, the bottleneck of frontier research shifts from human engineering headcount to compute orchestration. Startups with smaller teams can suddenly iterate experiments at the cadence of mega-labs.

The clear risk, however, lies in runaway blind spots. Handing over 100-round autonomous scaffold modifications means human oversight is reduced to glancing at final dashboards. If an evolving model optimizes for narrow metrics at the expense of system edge cases, diagnosing what went wrong three iterations deep could become an engineering nightmare.