MiniMax M2.5: Tiny, near Opus level, dirt cheap, fast AF
MiniMax M2.5 official release: a powerful, affordable domestic model with near Opus-level performance. Ultra-fast 100 tokens/sec, $0.3 per 1M input tokens.
“AI Disruption” Publication 8800 Subscriptions 20% Discount Offer Link.
MiniMax M2.5 has been officially released, and it’s fair to say this is an exceptionally strong Chinese model. Each domestic model has its own focus, and MiniMax is pursuing extreme optimization. Rather than flashy technical experiments, they are focused on delivering a model that is extremely easy to deploy, highly affordable, capable of getting work done, and competitive with the world’s top models in performance.
Let’s not overhype it—good-looking data doesn’t necessarily mean a great real-world experience—but at the very least, M2.5 is genuinely making an effort toward usability in real production environments.
Let me first share two images for you to get a sense of this.
The evolution of MiniMax - let’s look at the journey of the MiniMax M series:
Size comparison - this is interesting. It feels highly practical and very suitable for home lab deployment. It’s said that inference service providers might be able to extract amazing tokens-per-second generation speeds from this model.
