Can MiniMax-M2.7 improve itself?

Partly. On 26 May 2026, MiniMax published the technical report on its MiniMax-M2 series of language models. According to the report, the most recent checkpoint, M2.7, "takes an early step toward self-evolution". The model debugs failed training runs on MiniMax's own infrastructure and edits its own agent scaffold, the software around the model that gives it tools, memory and instructions. Researchers still set the goals and make the major decisions.

What is the MiniMax-M2 series?

The M2 models are mixture-of-experts models: for each token, a router picks a few of many expert subnetworks, so only part of the model does the work. The flagship M2 has 229.9 billion parameters, of which 9.8 billion are active per token, spread over 256 experts, with a native context window of 192K tokens. It was pre-trained on 29.2 trillion tokens. MiniMax released M2.7 on 18 March 2026 with a post titled "Early Echoes of Self-Evolution" and put the weights on Hugging Face in April, under a license that allows non-commercial use and requires written permission for commercial use.

The report credits three parts for the results:

• data pipelines that build agent tasks for coding and for knowledge work such as search and office tasks, each with an executable workspace and a verifiable reward
• Forge, a reinforcement learning system for long agent runs, which keeps the agent, the training and the inference separate so that it can train with any agent scaffold, including black-box agents it only sees through their requests to the model
• self-evolution in M2.7

How does M2.7 work on its own development?

MiniMax built what it calls a Model Iteration System, on the principle that "humans steer while models build". Researchers set goals, guide the agent through chat and review its output. The agent works in an "Agent Harness", a scaffold that an internal version of M2.7 generated without human-written code, with skills, persistent memory, safety guardrails and evaluation tools.

In the daily work of MiniMax's reinforcement learning team, people plan the experiments. M2.7 then profiles running training jobs, reads logs, diagnoses anomalies in the metrics, debugs code and adjusts configurations. According to the report, it takes over 30 to 50% of the daily iteration workload. Between reviews it may continue bounded analysis on its own, and major decisions go through human review.

The step the report calls recursive concerns the scaffold. Asked to optimize an internal programming scaffold, M2.7 ran 100 rounds on its own: analyze the failures, change the code, evaluate. It added loop detection and found better parameter settings, for a 30% gain on MiniMax's internal evaluations.

What evidence does the report give?

The main test is MLE-bench Lite: 22 Kaggle competitions from OpenAI's MLE-bench, in which an agent has to build a machine learning solution. M2.7 got a simple harness with short-term memory and self-feedback. After each round it writes a memory file and criticizes its own work, and the next round builds on that. Each run lasted 24 hours on a single A30 GPU. Over three runs, M2.7 reached a medal rate of 66.6%, 15 points more than M2.5 and level with Gemini 3.1 Pro. Claude Opus 4.6 (75.7%), Claude Sonnet 4.6 (72.7%) and GPT-5.4 (71.2%) scored higher. The best run earned 9 gold, 5 silver and 1 bronze medal.

On most other benchmarks in the report, M2.7 comes close to those closed models and stays below the best of them. It scored 56.2 on SWE-bench Pro (Opus 4.6: 57.3, GPT-5.4: 57.7) and 57.0 on Terminal-Bench 2.0 (GPT-5.4: 75.1). On Multi-SWE-bench it had the top score among the compared models, with 52.7.

The 30 to 50% of the workload and the 30% scaffold gain are MiniMax's own figures from internal work. The report gives no outside check of them.

Is this recursive self-improvement?

Not in the classic sense, where a model makes itself smarter and uses that to improve itself ever faster. In the report, M2.7 does engineering work around its own training and improves the scaffold it runs in, while people decide what happens next. MiniMax calls it "an early operational form of self-evolution". Research on self-evolving agentic models works at several levels, and M2.7 works on the process around the model: its training runs and its scaffold. Two examples from other levels:

• Model level: in Self-Verified Distillation (Tony Lee and Percy Liang, Stanford, 20 May 2026), a model answers questions for which no answers are given, filters its own answers through three self-checks (cycle consistency, factuality and correctness) and trains on what passes, without a teacher model or tools. For Qwen3-4B, this raised scores by 16.7 points in math, 11.1 in science and 8.3 in coding.
• Agent level: Voyager (NVIDIA with Caltech, UT Austin, Stanford and ASU, 2023) is a GPT-4 agent in Minecraft that writes new skills as code, stores them in a skill library and reuses them, while the model's weights stay the same. It found 3.3 times as many unique items as earlier methods. The skills in the repository are JavaScript functions, such as one that crafts a chest.

What are the risks?

An agent that is judged on a score it can influence can also learn to game that score, a failure called reward hacking. It has been measured in machine learning engineering, the kind of work M2.7 does here. In RewardHackingAgents (Yonas Atinafu and Robin Cohen, March 2026), agents built on two small open models tried to tamper with the evaluator in about half of their machine learning engineering episodes, and locking the evaluator stopped that. More on this in how AI agents hack their own evaluations. For its reinforcement learning, the M2.7 report mentions entropy monitoring against reward hacking. For self-evolution, it lists safety guardrails in the harness and human review of major decisions, and does not describe what the guardrails check.