The official results show M2.7 covering a broad range of real engineering, end-to-end project, and Office/Agent tasks: 56.22% on SWE-Pro, 55.6% on VIBE-Pro, 57.0% on Terminal Bench 2, 1495 on GDPval-AA, and 62.7% on MM Claw, although the complete harness remains undisclosed.
Model: MiniMax M2.7; official API/Agent and internal Agent harness.
Engineering benchmarks: SWE-Pro, VIBE-Pro, Terminal Bench 2, NL2Repo, SWE Multilingual, and Multi SWE Bench.
Office/Agent benchmarks: GDPval-AA, Toolathlon, MM Claw (40+ skills/complex work), and MLE Bench Lite (22 ML competitions).
Configuration: Each benchmark was run by an official/internal harness; the article does not provide the complete inputs, sampling settings, repeat counts, costs, or confidence intervals for each test.
The official cases include debugging workflows involving production logs/monitoring/deployment timelines/traces/databases/codebases, delivery of Web/Android/iOS/simulated projects, multi-round Word/Excel/PPT editing, research reports, and a TSMC revenue model. Some results came from internal complex skills, Agent Teams, and dynamic tool search.
SWE-Pro: 56.22%; SWE Multilingual: 76.5; Multi SWE Bench: 52.7.
VIBE-Pro: 55.6%; Terminal Bench 2: 57.0%; NL2Repo: 39.8%.
GDPval-AA: ELO 1495, described by the official source as the highest among open-source models; Toolathlon: 46.3%.
MM Claw: 62.7%, described by the official source as close to Sonnet 4.6; across 40+ complex skills, each exceeding 2,000 tokens, skill adherence was 97%.
MLE Bench Lite: average medal rate of 66.6% across three 24-hour trials; the best run achieved 9 gold/5 silver/1 bronze.
M2.7 is suitable for Agents that need real-world engineering-system understanding, end-to-end delivery, and tool/skill orchestration. When using the public API, however, treat the official internal harness as an upper-bound reference and remeasure performance with the base model or your own scaffold.
These are vendor-published benchmarks; the complete prompts, task splits, harnesses, costs, failure samples, and statistical variance are undisclosed.
Many figures depend on MiniMax's internal Agent Teams, memory, skills, and dynamic tool search, and cannot be attributed directly to the base model.
GDPval-AA, MM Claw, SWE-Pro, and other metrics have different properties and should not be combined into one overall score.
“reduced recovery time under three minutes” refers to multiple internal cases reported by the official source, not controlled average performance.
Clearly choose the API, open weights, vLLM/SGLang, or a custom Agent harness, and lock the version.
Select coding issues, end-to-end projects, Terminal, office-document, tool-calling, and long-running experiment tasks, and provide the same tools and permissions.
Record prompts, skill/tool definitions, call traces, patches, tests, costs, latency, and human intervention.
Report results separately by benchmark and harness, distinguishing base-model performance from gains due to Agent orchestration.
MiniMax M2.7
