Minimax M2.7 is INSANELY GOOD! (Full Review)

M2.7 is a post-trained M2.5 — same 230B MoE, sharper agent behavior

Minimax M2.7 is not a new base model. It keeps M2.5's architecture: 230B total parameters, mixture-of-experts, 10B active per token. The jump comes from continued post-training that MiniMax describes as "beginning the journey of recursive self-improvement." On the creator's Discord, the seven OpenClaw agents running Minimax were switched to M2.7 via the same API key and endpoint — just update the model parameter (or run /status in Discord to confirm).

Three areas MiniMax officially targeted

MiniMax lists three improvements over M2.5: real-world engineering (day-to-day coding, not just benchmarks), professional office delivery (documents, analysis, business tasks — a new focus for 2.7), and "character-rich interaction" with more personality in conversation. In a one-shot presentation test, M2.7's agent (Gambit) spawned parallel sub-agents for research, presentation, and self-audit, while M2.5's agent did everything solo and produced a slide that flashed OpenClaw 3.7 on every transition.

Where 2.7 actually moves the needle

Early testing shows M2.7 is less prone to overthinking on complex tasks, with tighter instruction following and improved tool calling — M2.5 already scored 76.8% on BFCL, and 2.7 should push that further. M2.5 already beat Claude Opus 4.6 and Sonnet 4.6 on multi-SWE-bench and BFCL multi-turn; M2.7 is expected to match or exceed those numbers once published. Speed held up: a one-shot presentation finished in 2–3 minutes, compared to 10–15 minutes for Claude on the same task. Raw knowledge is still a weak spot — the model is positioned as an agentic coding model, not a general chatbot.

Routing strategy the creator actually uses

Stark (orchestrator) stays on Opus 4.6 for deep reasoning and autonomous terminal work; the research and specialized agents stay on Minimax 2.7. Use 2.7 for iterative coding, multi-file refactors, long agentic loops, and Go/Rust/TypeScript/Java work (Minimax's reported edge). Use Opus 4.6 for deep reasoning, terminal ops, and architecture setup. Use GPT 5.4 or Notebook LM for planning — the creator calls Notebook LM "really good for planning compared to building from scratch with Claude code." Note: the high-speed variant requires the $40/month tier; the rest can stay on the plus plan.