Short answer: MiniMax’s March 18, 2026 release announcement reports M2.7 scores of 56.22% on SWE-Pro, 55.6% on VIBE-Pro, 57.0% on Terminal Bench 2, and 1,495 ELO on GDPval-AA. These are vendor-reported results, not independent evaluations. The announcement does not provide enough test detail to establish a like-for-like ranking against other models. (minimax.io)
There’s also an important API caveat: as of November 10, 2026, M2.7 is close to retirement on MiniMax’s API platform. New users lost access to M2.7 on October 19. Existing pay-as-you-go API users can use it until its scheduled retirement on November 18, 2026. M Plan and Token Plan requests using M2.7 names have been routed to M3.1 Flash since October 19, so keeping the old model name does not mean those requests still run M2.7. (platform.minimax.io)
MiniMax M2.7 benchmark scores
| Evaluation | MiniMax-reported result | What the announcement says | What it does not establish |
|---|---|---|---|
| SWE-Pro | 56.22% | Software engineering performance; MiniMax says the result nearly approaches Opus’s best level. | The score definition, evaluation setup, named Opus version, and comparable test conditions. |
| VIBE-Pro | 55.6% | Performance in end-to-end project delivery scenarios. | The task set, scoring procedure, harness, and an independently verified comparison. |
| Terminal Bench 2 | 57.0% | Performance on understanding complex engineering systems. | The exact test configuration, tools, scoring details, or comparative baseline. |
| GDPval-AA | 1,495 ELO | Professional office-software tasks; MiniMax describes it as the highest score among open-source models. | The leaderboard snapshot, comparison set, and evaluation conditions behind that claim. |
All four figures and MiniMax’s descriptions come from the company’s M2.7 announcement. The page gives scores, but the reported material does not spell out benchmark protocols or enough conditions to independently reproduce the results. Treat them as useful signals about what MiniMax chose to report—not as a complete, independently verified leaderboard. (minimax.io)
MiniMax also says M2.7 achieved a 97% skill-adherence rate while working with more than 40 complex skills, each exceeding 2,000 tokens. That is a separate company-reported capability claim; the announcement does not provide a test protocol that would show how to compare it directly with another model. (minimax.io)
What the scores mean for model selection
The reported benchmarks point toward software engineering, agent-style task execution, and office productivity as areas MiniMax evaluated. They do not show that M2.7 will be best for every coding task, tool chain, or business workflow. Nor do the available details support a fair numerical comparison with a named competitor: that requires matching the model versions, benchmark, scoring method, and test conditions.
For a coding-agent evaluation, for example, SWE-Pro or Terminal Bench 2 may be relevant signals. But if your production tasks involve a particular language, repository size, tool setup, or review policy, the benchmark result alone cannot tell you how often the model will succeed in that environment. A practical selection test is to run the candidate model on a representative set of your own tasks, use the same prompt, tools, and success criteria for every model, and record task completion and any operational measures that matter to your team. That is a proposed evaluation method, not a claim that such a test was performed here.
Decision: Consider the published scores a reason to investigate M2.7’s fit for a matching task—not proof of a universal ranking or production readiness. The evidence is MiniMax-reported, and the released announcement does not provide enough methodology to settle close comparisons.
Is M2.7 still a production API option?
That depends on your MiniMax access route and date:
- New MiniMax API users: M2.7 stopped being available to new users on October 19, 2026.
- Existing pay-as-you-go API users: MiniMax says access is scheduled to continue until November 18, 2026, after which requests to M2.7 will no longer work.
- M Plan and Token Plan users: Since October 19, 2026, requests using M2.7 model names are automatically routed to M3.1 Flash. The name can remain in a request for backward compatibility, but the response is not evidence of M2.7 performance.
- Self-hosting: MiniMax’s documentation provides a deployment path for the M2.7 weights using SGLang, with a reference setup using four high-memory NVIDIA GPUs. The same documentation says commercial use has attribution and prior-authorization requirements; review the complete license before production or commercial use. This is a separate route from the MiniMax-hosted API and has its own infrastructure and licensing considerations. (
platform.minimax.io)
So if you need MiniMax’s hosted API now, don’t select or benchmark M2.7 by model name alone. Confirm which model your account and plan actually serve. If you need M2.7 specifically, the documented self-hosted path is an option to investigate, subject to hardware, deployment, and license requirements. For a new hosted-model evaluation, MiniMax’s current model documentation lists M3 and M3.1 Flash as available language models; their performance should be tested separately rather than inferred from M2.7’s scores. (platform.minimax.io)
Bottom line
MiniMax reports 56.22% on SWE-Pro, 55.6% on VIBE-Pro, 57.0% on Terminal Bench 2, and 1,495 ELO on GDPval-AA for M2.7. Those figures make the release announcement relevant to benchmark-focused research, but they are vendor-reported and do not include enough methodology for robust cross-model ranking. More immediately, M2.7’s hosted API lifecycle is ending: new users lost access on October 19, and existing pay-as-you-go API access is scheduled to end November 18, 2026. Choose it only if your access route genuinely serves M2.7, and validate it against your own workload; otherwise evaluate the currently available model or a properly licensed self-hosted deployment.
Sources
MiniMax M2.7: Early Echoes of Self-Evolution