MiniMax M2.7: Early Echoes of Self-Evolution

TL;DR

MiniMax M2.7 is the next iteration in the M2 series (after M2 → M2.1 → M2.5), positioned around a single distinguishing claim: it is “MiniMax’s first model deeply participating in its own evolution.” The blog reports SWE-Pro 56.22, VIBE-Pro 55.6, Terminal Bench 2 57.0, GDPval-AA ELO 1495 (highest among open-source models), Toolathon 46.3, and a 97% skill-adherence rate across 40 complex >2K-token skills — plus a striking MLE-Bench Lite result where a simple short-term-memory + self-feedback + self-optimization harness drives M2.7 to a 66.6% average medal rate over three 24-hour runs (best run: 9 gold / 5 silver / 1 bronze across 22 competitions), second only to Opus-4.6 and GPT-5.4. The most concrete self-evolution datapoint is a 100-round autonomous scaffold-optimization loop (“analyze failure trajectories → plan changes → modify scaffold code → run evaluations → compare results → decide to keep or revert”) that yielded a 30% performance improvement on internal evals. An open-source companion demo (OpenRoom) ships alongside the API/Agent release. No technical report yet — recipe specifics, architectural changes from M2.5, and the full benchmark suite are not disclosed.

Key claims

M2.7 reports SWE-Pro 56.22%, matching GPT-5.3-Codex and “nearly approaching Opus’s best level”; VIBE-Pro 55.6% (“nearly on par with Opus 4.6”); Terminal Bench 2 57.0%; NL2Repo 39.8%; SWE Multilingual 76.5; Multi SWE Bench 52.7 [§Professional Software Engineering, §1].
GDPval-AA ELO 1495 is reported as the highest among open-source models (vs Opus 4.6, Sonnet 4.6, GPT5.4 ahead; surpassing GPT5.3) across 45 ranked models [§Professional Work, §2].
M2.7 maintains a 97% skill-adherence rate across 40 complex skills each exceeding 2,000 tokens (MM Claw setup), and reaches 46.3% on Toolathon — described as “global top tier” [§Professional Work, §2].
An internal research-agent harness built by M2.7 itself handles 30–50% of the daily RL-team workflow (literature review, experiment tracking, data piping, log reading, debugging, metric analysis, code fixes, merge requests, smoke tests), with human researchers reserved for critical decisions [§Building an agent for model self-evolution].
The concrete self-evolution datapoint: M2.7 autonomously ran a 100-round loop on an internal scaffold (“analyze failure trajectories → plan changes → modify scaffold code → run evaluations → compare results → decide to keep or revert”), discovering three categories of optimization — (a) sampling-parameter combinations (temperature, frequency penalty, presence penalty); (b) more-specific workflow guidelines (e.g. automatically searching for the same bug pattern in other files after a fix); (c) loop detection and other agent-loop optimizations — yielding a 30% performance improvement on internal evaluation sets [§Building an agent for model self-evolution].
On MLE Bench Lite (22 single-A30-GPU competitions) under a simple short-term-memory + self-criticism + self-optimization harness, M2.7 achieved an average medal rate of 66.6% across three 24-hour runs; best run: 9 gold / 5 silver / 1 bronze. Ranked behind Opus-4.6 (75.7%) and GPT-5.4 (71.2%), tied with Gemini-3.1 (66.6%) [§Building an agent for model self-evolution].
An MM Claw evaluation set (covering personal learning planning, office document processing, research/investment advice, and code development/maintenance) reports M2.7 at 62.7% — “close to Sonnet 4.6” [§Professional Work].
M2.7’s recovery time for live-production-system debugging incidents (alert correlation → trace statistical analysis → causal hypothesis → database verification → non-blocking index migration MR) has on multiple occasions been reduced to under three minutes [§Professional Software Engineering, debugging vignette].
OpenRoom — a web-GUI interaction system for character-consistent agents — is open-sourced alongside the release (github.com/MiniMax-AI/OpenRoom), with the blog noting that “most of the code was written by AI” [§Entertainment].

Method

The blog discloses no architectural changes from M2.5 and no training recipe; the entire technical disclosure is at the harness and workflow level, not the model. The interesting axis is the research-agent harness MiniMax built around M2.7 itself.

The harness supports data pipelines, training environments, infrastructure, cross-team collaboration, and persistent memory. The exemplary RL-team workflow chains: idea discussion with the agent → literature review → experiment-spec tracking → data/artifact piping → experiment launch → automatic monitoring and profiling → triggered log reading, debugging, metric analysis, code fixes, merge requests, and smoke tests. Critically, the harness itself is recursively evolved by M2.7: it autonomously collects feedback, builds internal-task evaluation sets, and iterates its own architecture, skills/MCP implementation, and memory mechanisms.

The 100-round scaffold-optimization loop is the cleanest documented instance: M2.7 was tasked with improving a model’s programming performance on an internal scaffold and ran an autonomous outer loop of analyze → plan → modify → evaluate → compare → keep/revert for over 100 rounds. The discovered optimizations split into three categories: sampling-parameter sweeps; more-specific workflow guidelines (e.g. propagating bug-fix patterns across other files); and agent-loop optimizations like loop detection. The reported outcome is a 30% improvement on internal evaluation sets.

The MLE Bench Lite experiment is a lower-resource self-improvement probe: a minimal harness with three components — short-term memory (markdown file regenerated per round), self-feedback (per-round self-criticism), and self-optimization (next round conditions on the memory + feedback chain from all prior rounds). Three independent 24-hour trials were run across 22 competitions runnable on a single A30 GPU.

The Agent Teams setup (multi-agent collaboration with native role differentiation, adversarial reasoning, protocol adherence, and behavioral differentiation) is framed as a paradigm that “cannot be achieved through prompting alone and must be internalized as native capabilities of the model” — i.e. baked into post-training rather than scaffolded externally.

Results

Software engineering (the headline axis):

SWE-Pro 56.22%, Terminal Bench 2 57.0%, NL2Repo 39.8%, VIBE-Pro 55.6%, SWE Multilingual 76.5, Multi SWE Bench 52.7.

Professional work:

GDPval-AA ELO 1495 (highest open-source, ahead of GPT5.3, behind Opus 4.6 / Sonnet 4.6 / GPT5.4).
Toolathon 46.3%; MM Claw 62.7%; 97% skill adherence across 40 × >2K-token skills.

Self-evolution:

100-round autonomous scaffold-optimization loop → 30% gain on internal evals.
MLE Bench Lite: 66.6% average medal rate (3 × 24h trials, 22 competitions); best run 9 gold / 5 silver / 1 bronze.
Internal research-agent: M2.7 handles 30–50% of the daily RL-team workflow.

No public benchmarks reported for HLE, AIME, GPQA, BrowseComp, BFCL multi-turn, or SWE-Bench Verified (the previous M2.5 headline metric). No model card, technical report, or cost/throughput numbers in the blog. The benchmark set is qualitatively different from M2.5’s filing — M2.7 is positioned around SWE-Pro / Terminal Bench 2 / GDPval-AA / Toolathon, not SWE-Bench Verified + BrowseComp + BFCL — making direct M2.5 → M2.7 deltas hard to compute from public material.

Why it’s interesting

This is the next step in the MiniMax open-frontier-model curve already tracked at Luma via MiniMax-M2.5, but the framing has shifted axis: M2.5 was about agentic RL scaling produces a frontier coding agent at 1/10–1/20 the cost, M2.7 is about the coding agent now closes the loop and improves its own training/serving infrastructure. The 100-round autonomous scaffold-optimization loop is the cleanest filed instance of a frontier LLM doing what GPT-5.3-Codex’s release blog described abstractly as “first model instrumental in creating itself” (Introducing GPT-5.3-Codex) — except here the optimization target (model serving harness on an internal eval set) and the discovered optimizations (sampling-parameter sweeps, cross-file bug-pattern propagation, loop detection) are spelled out.

The self-evolution claim also gives the wiki a second concrete datapoint adjacent to the In-Situ Self-Evolving paradigm of Yunjue Agent Tech Report: A Fully Reproducible, Zero-Start In-Situ Self-Evolving Agent System for Open-Ended Tasks: Yunjue evolves a tool pool under binary execution feedback from a zero-tool start; M2.7 evolves a scaffold/harness under aggregate-eval-score feedback. Both reduce supervision to verifiable signals (code runs / score improves) and use the model itself as the optimization driver. The mechanism differs (Yunjue’s Parallel Batch Evolution synthesizes tools; M2.7’s outer loop modifies scaffold code and decides to keep or revert) but the load-bearing primitive is the same.

The MLE Bench Lite 66.6% medal rate is the third interesting datapoint — second only to Opus-4.6 and GPT-5.4 across 22 ML competitions under a simple 24-hour memory+self-feedback harness. Worth tracking alongside Scaling Karpathy's Autoresearch: What Happens When the Agent Gets a GPU Cluster (Karpathy-style autoresearch with a GPU cluster), FARS: Fully Automated Research System (Fully Automated Research System), and the broader autoresearch trend: this is the first filed datapoint where the base model (not an agent system with hand-built scaffolds) is the variable, with the harness held fixed and minimal.

What’s missing: no technical report, no architecture/recipe disclosure vs M2.5, no SWE-Bench Verified or BrowseComp / BFCL numbers, no cost/throughput envelope. The benchmark axis has shifted to SWE-Pro + Terminal Bench 2 + GDPval-AA, making like-for-like comparison against the M2.5 filing hard. The 30% scaffold-optimization gain is on an internal eval set; no external probe of whether the discovered optimizations transfer.