MiniMax · 2026-03-18
MiniMax M2.7: Early Echoes of Self-Evolution
“M2.7 is our first model deeply participating in its own evolution.”
- Who
- MiniMax
- Date
- 2026-03-18 The post's text shows no date; the date comes only from the reader's published-time metadata for the page.
- Evidence
- Not rated. A description of how the model was built. The 30% figure is on internal evaluation sets the post does not describe. how we rate evidence
- Replication
- Nothing to rerun. all claims
- Checked
- 2026-10-29, against the source page
- Links
- Source
Also on the page
- On M2.7's development, MiniMax says "we let the model update its own memory and build dozens of complex skills in its harness to help with reinforcement learning experiments". In a separate section, an internal harness built by "an internal version of M2.7" "continuously iterates its own architecture, skills/MCP implementation, and memory mechanisms".
- One example: "we had M2.7 optimize a model's programming performance on an internal scaffold", in a loop run "for over 100 rounds", which "achieved a 30% performance improvement on internal evaluation sets".
- In what MiniMax calls "preliminary exploratory tests in low-resource scenarios", with "a simple harness", M2.7 took part in 22 competitions at the MLE Bench Lite level, three trials of 24 hours each: "The average medal rate across the three runs was 66.6%".
- MiniMax on what comes next: "We believe that future AI self-evolution will gradually transition towards full autonomy, coordinating data construction, model training, inference architecture, evaluation, and other stages without human involvement."