CONFIRMEDDate confidence: CONFIRMEDStated as firmHow firmly the source words the date: confirmed, scheduled, estimated, delayed or unverified.Show all CONFIRMED pins87% CONFIDENCEOverall confidence: 87%How well the pin's source and references back up its dates.Weighted average of how firmly 6 references, the source included, support the pin's start and end times; a reference counts half as much for every 180 days older than the newestShow all pins at 75% confidence or better
MiniMax's[1][3] post "MiniMax M2.7: Early Echoes of Self-Evolution" is dated 2026-03-18 and says M2.7 is "now fully available on MiniMax Agent and the MiniMax API Platform"; the API release notes list it on Mar. 18, 2026.- Start
- Mar 18, 202690% CONFIDENCEfrom the source
MiniMax-M2.7 Released
MiniMax releases M2.7, its first model to take part in its own evolution, scoring 56.22% on SWE-Pro and 57.0% on Terminal Bench 2 with native Agent Teams, at $0.30 per 1 million input tokens.
- MiniMax released M2.7 on 18 March 2026 as "our first model deeply participating in its own evolution", available the same day on MiniMax Agent and the API[1]
- Self-evolution is the headline: an internal M2.7 ran a loop of analysing failure trajectories, planning changes, editing its own scaffold code and re-running evaluations for more than 100 rounds, for roughly a 30% gain on internal sets, and MiniMax's RL team reportedly has it handling 30% to 50% of daily workflows end to end[1][6]
- On coding it reaches 56.22% on SWE-Pro (MiniMax says it matches GPT-5.3-Codex), 55.6% on VIBE-Pro, 57.0% on Terminal Bench 2, 76.5 on SWE Multilingual and 52.7 on Multi SWE Bench[1][2]
- For office work it posts a GDPval-AA Elo of 1495, the highest among open models and behind only Opus 4.6, Sonnet 4.6 and GPT 5.4 on that board, and keeps 97% skill adherence across more than 40 skills of over 2,000 tokens each[1]
- The weights reached Hugging Face in April[6]; the API charges $0.30 per 1 million input tokens and $1.20 for output, with a highspeed tier at double that[3]; Artificial Analysis scores it 23 on its current index, with a 205k-token context[4]
- M3 replaced it on 1 June with a 1M-token context and image and video input, which M2.7 lacks as a text-only model[4]
Notable features
- Native Agent Teams: role boundaries, adversarial reasoning and protocol adherence are built into the model rather than prompted, and MiniMax shows a live-production debugging demo that correlates monitoring metrics with deployment timelines[1]
- MLE Bench Lite: a 66.6% average medal rate over three 24-hour trials (9 gold, 5 silver and 1 bronze in the best run), against 75.7% for Opus 4.6, 71.2% for GPT 5.4 and 66.6% for Gemini 3.1 Pro[6]
- Character consistency and emotional intelligence were improved, with OpenRoom, an interactive agent-harness demo, as the showcase[1]
- It takes text only and has 229B total and about 10B active parameters, the same shape as M2 and M2.5[2][4][5]
Benchmarks[1]
| Benchmark | M2.7 | M2.5 | Gemini 3.1 Pro | Sonnet 4.6 | Opus 4.6 | GPT 5.4 |
|---|---|---|---|---|---|---|
| SWE-Bench Pro | 56.2 | 55.4 | 54.2 | 57.2 | 57.3 | 57.7 |
| Multi-SWE-Bench | 52.7 | 51.3 | — | 51 | 50.3 | 49 |
| VIBE-Pro | 55.6 | 54.2 | 41 | 56.1 | 55.6 | — |
| MLE-Bench lite (medal rate) | 66.6 | 51.5 | 66.6 | 72.7 | 75.7 | 71.2 |
| Toolathlon | 46.3 | 38.3 | 48.8 | 44.8 | 47.2 | 54.6 |
| MM-ClawBench | 62.7 | 57.6 | 61.8 | 64.2 | 75.4 | 73.6 |
| Artificial Analysis index (at launch) | 50 | 42 | 57 | 52 | 53 | 57 |
- Read from MiniMax's benchmark chart on the launch post; GDPval-AA is quoted in the text as an Elo of 1495, so it is not in the table[1]