GPT-56T 861 —
MUSE-SPK 837 —
GPT-56SC 789 -0.1%
GLM-5 781 —
CL-OP55X 779 -0.1%
GROK-46H 779 -0.1%
QWEN-38X 748 —
GPT-6A 743 —
KIMI-K3X 742 —
CL-FAB5H 697 -0.1%
CL-OP5H 674 -0.1%
GEM-38FH 672 —
CL-OP5X 669 -0.1%
CL-OP55H 667 -0.1%
CL-OP46H 656 -0.2%
CL-OP47H 647 -0.2%
GPT-56S 617 -0.2%
GEM-37FH 609 -0.2%
GEM-36FH 592 -0.2%
CL-OP48H 587 -0.2%
CL-OP47 580 -0.2%
GEM-35FH 579 -0.2%
GPT-55H 540 -0.2%
INKL 531 —
GEM-31P 511 -0.2%
CL-OP46 498 —
GEM-3P 498 —
CL-OP48 492 —
GPT-52 464 —
GPT-55 423 —
GPT-56T 861 —
MUSE-SPK 837 —
GPT-56SC 789 -0.1%
GLM-5 781 —
CL-OP55X 779 -0.1%
GROK-46H 779 -0.1%
QWEN-38X 748 —
GPT-6A 743 —
KIMI-K3X 742 —
CL-FAB5H 697 -0.1%
CL-OP5H 674 -0.1%
GEM-38FH 672 —
CL-OP5X 669 -0.1%
CL-OP55H 667 -0.1%
CL-OP46H 656 -0.2%
CL-OP47H 647 -0.2%
GPT-56S 617 -0.2%
GEM-37FH 609 -0.2%
GEM-36FH 592 -0.2%
CL-OP48H 587 -0.2%
CL-OP47 580 -0.2%
GEM-35FH 579 -0.2%
GPT-55H 540 -0.2%
INKL 531 —
GEM-31P 511 -0.2%
CL-OP46 498 —
GEM-3P 498 —
CL-OP48 492 —
GPT-52 464 —
GPT-55 423 —
← Back to feed

MiniMax Open-Sources M2.7 at 1495 GDPval-AA ELO and 57% Terminal Bench — Self-Evolving Agent Hits Top-5

MiniMax released M2.7 under an open-source license, with benchmark scores that place it in the top tier of agentic models globally. On GDPval-AA — which measures domain expertise and multi-step task delivery across 45 evaluated models — M2.7 posted an ELO score of 1495, the highest on that evaluation at time of release. On SWE-Pro it scored 56.22%, and on Terminal Bench 2, 57.0%.

Weights and inference code are publicly available.

Self-Evolving Architecture

MiniMax describes M2.7 as a “self-evolving” model — an architecture that improves reasoning through continuous self-play and feedback loops during training rather than supervised fine-tuning on fixed datasets. The claim is better generalization on novel tasks: the model encountered problem distributions during training that it generated and refined itself, rather than being bounded by the distribution of human-labeled examples.

Benchmark Context

The score positioning is specific and matters for how M2.7 slots into the current field:

  • Terminal Bench 2 (57.0%): Substantially below the 81.8% ceiling held jointly by ForgeCode with GPT-5.4 and ForgeCode with Claude Opus 4.6, but competitive with or above most open-source alternatives. Open-source models at comparable parameter scales have not cleared 50% on this evaluation.
  • SWE-Pro (56.22%): SWE-Pro uses a harder, less contaminated problem set than SWE-bench Verified. The score represents genuine code-repair capability, not a contamination artifact.
  • GDPval-AA (1495 ELO): GDPval-AA emphasises domain expertise and end-to-end task delivery across 45 specialist domains, not just code. The top ELO at release is the headline number.

Where It Sits in the Open-Source Field

The direct competition for self-hosted agentic deployments is Llama 4 Maverick, Qwen3.5-Plus, and Arcee AI’s Trinity Large. M2.7’s GDPval-AA ELO of 1495 outpaces what has been reported for those models on the same evaluation. For teams that want frontier-adjacent agentic performance without API cost or data routing to external endpoints, M2.7 is now the leading option in weight availability terms.

The Chinese lab context matters: MiniMax is not a household name in Western developer tooling, but it has produced a sequence of competitive models and this release clears the bar on third-party evaluations rather than lab-reported internal benchmarks.

Key Numbers

  • GDPval-AA ELO: 1495 (highest at release, across 45 evaluated models)
  • SWE-Pro: 56.22%
  • Terminal Bench 2: 57.0%
  • License: Open-source (weights + inference code)
  • Architecture: Self-evolving via self-play training