A new deployment tier for engineering, tool use, and office productivity
MiniMax M2-compatible APIs stay the same; migration usually means changing only the model value.
Overview
MiniMax M2.7 is MiniMax's newer text model for real engineering work, tool use, search, and office productivity. MiniMax describes M2.7 as reaching or refreshing SOTA in programming, tool calling, search, and office productivity, while exposing both OpenAI-compatible Chat Completions and Anthropic-compatible Messages APIs.
Compared with M2.5, M2.7 is a new deployment tier: same request and response protocols, but a separate model ID, separate usage attribution, and independent rollout control. Existing MiniMax M2 integrations can usually move to M2.7 by changing the model value.
Key capabilities
| Dimension | Detail |
|---|---|
| Context window | 204,800 tokens |
| Max output | 2,048 tokens |
| Input modalities | Text |
| Output modalities | Text |
| Tools | streaming, tool use, interleaved thinking, OpenAI / Anthropic compatible API |
MiniMax recommends the Anthropic-compatible Messages API as the default path for thinking blocks, interleaved thinking, and other advanced capabilities. See live pricing in the model catalog.
Software engineering
Three core scores for real engineering work
The focus is end-to-end projects, system understanding, log analysis, and production debugging.
MiniMax's announcement focuses on real engineering tasks. M2.7 scores 56.22% on SWE-Pro, 55.6% on VIBE-Pro across Web, Android, iOS, and simulation-style end-to-end projects, and 57.0% on Terminal Bench 2. These benchmarks emphasize repository context, logs, environments, and toolchains rather than short code snippets.
Self-evolution and office productivity
From research agents to editable office deliverables
The important shift is from one-shot text answers to deliverables that can be revised across rounds.
MiniMax highlights "self-evolution": internally, M2.7 drives a research-agent harness that reads papers, tracks experiment specs, starts experiments, monitors logs, analyzes metrics, edits code, and launches smoke tests. MiniMax says it can handle roughly 30%-50% of daily RL-team workflows. Office benchmarks include GDPval-AA 1495 ELO, Toolathon 46.3%, and 97% skill adherence across 40 MM Claw skills.
When to use it
- Real software engineering: end-to-end implementation, debugging, logs, and security tasks.
- Code plus tools: shell, browser, search, file editing, and iterative validation.
- Office automation: Excel, PowerPoint, Word, financial modeling, and business documents.
- Smooth migration: swap the model ID while keeping OpenAI / Anthropic-compatible APIs.
CrossModel exposes MiniMax M2.7 through OpenAI-compatible /v1/chat/completions and Anthropic-compatible /v1/messages. Current pricing is available in the model catalog.