What MiniMax M2.7 Is
MiniMax M2.7 is MiniMax's agent-focused model — a 230B MoE (~10B active) explicitly trained for multi-agent collaboration, live debugging, and end-to-end delivery across multi-step pipelines. It sits in the open-weight value sweet spot: a 50 on the Artificial Analysis Intelligence Index (close to closed-source flagships) at roughly a tenth of their price, with a ~205K context window and native Agent Teams / dynamic tool search.
Pricing
$0.30 input / $1.20 output per 1M tokens — among the cheapest capable agent models. A HighSpeed variant runs ~2x the price for roughly double the throughput (~100 vs ~48 tok/s). On LinkModel it runs behind one key alongside Claude, GPT, Gemini, DeepSeek, GLM, and Kimi, up to 30% below official.
When to Use It
- Long-horizon agents — research assistants, multi-step debugging, financial-modeling pipelines where the model must hold a coherent plan across dozens of tool calls. Its agent training shows up most on runs longer than ~5 minutes of continuous reasoning.
- Cost-sensitive agent workloads — near-flagship intelligence at value-tier price.
- SRE / office automation / ML research delivery — the use cases MiniMax tuned it for (MLE-Bench-Lite ~66.6%, trailing only the very top flagships).
When Not To
- Latency-critical streaming to users → standard M2.7 is ~48 tok/s; consider the HighSpeed variant or a faster model.
- Cheapest possible bulk text → DeepSeek V4 Flash ($0.14/$0.28) edges it on raw price.
- Multimodal input → M2.7 is text-focused; use Kimi K2.6 or Gemini.
Licensing Caveat
M2.7 ships under a Modified-MIT license that bars commercial self-hosting without authorization (unlike the fully-MIT M2/M2.5). For commercial use, route through a paid API or get separate authorization. Using it through a gateway sidesteps the self-hosting license question and adds zero data retention by default.
How to Call It
MiniMax is OpenAI-compatible. Through a gateway:
curl -X POST https://api.linkmodel.ai/api/v1/... \
-H "Authorization: Bearer $LINKMODEL_API_KEY" -H "Content-Type: application/json" \
-d '{ "model": "minimax-m2.7", "messages": [{"role":"user","content":"Plan and execute: migrate this service to async, step by step."}] }'Confirm the exact chat endpoint and schema in the docs.
Cost Control
- Route down — use M2.7 for the long agent plans, a cheaper model for simple steps. It's often itself the "cheap agent default" you route to. See how much it costs to run an AI agent.
- Cap output and prune context across the many calls an agent makes. More in how to reduce AI API costs.
Bottom Line
MiniMax M2.7 is one of the best value picks for long-horizon agents — near-flagship intelligence at ~$0.30/$1.20, tuned for multi-agent, multi-step delivery. Mind the Modified-MIT license for self-hosting. Compare in best LLM API and best coding LLM API.
Start free with a $1 credit and put it on a real agent task.