Question 1

Is MiniMax M2.7 free?

Accepted Answer

MiniMax M2.7 is not free to use through the API, costing $0.30 per million input and $1.20 per million output tokens [1]. The weights are downloadable at no cost from Hugging Face, but you pay for the hardware to run them yourself [7].

Question 2

How big is it?

Accepted Answer

M2.7 holds 230B total parameters with roughly 10B active per forward pass, using a mixture-of-experts architecture [4]. That sparsity keeps inference cost and latency low relative to the total parameter count.

Question 3

Is it better than Claude for coding?

Accepted Answer

It depends on the task. M2.7 matches Sonnet 4.6 on MM Claw and nears Opus 4.6 on VIBE-Pro, but Opus still leads on the hardest problems [3]. M2.7 wins decisively on price.

Question 4

Can I run it locally?

Accepted Answer

Yes. Download the weights from Hugging Face or ModelScope and serve them with SGLang, vLLM, or Transformers [7]. You need substantial GPU memory even with 10B active parameters.

Question 5

What is the difference between M2.7 and M2.7-highspeed?

Accepted Answer

Both produce identical results, but highspeed runs faster and costs double on output, $2.40 versus $1.20 per million tokens [1][3]. Choose highspeed for latency-sensitive work, standard for cost-sensitive volume.

Question 6

What changed from M2.5 to M2.7?

Accepted Answer

M2.7 added the self-evolution training loop, Agent Teams for multi-agent work, and stronger office editing, while shifting its headline coding benchmark to the harder SWE-Pro test at 56.22% [3].