What is MiniMax?

RL agents that appear to generalize via tool scaling actually overfit to the scaffold, failing when even small changes are made to the agent environment.

“We soon realized that if we perturb the environment a little bit, for example, we change another agent scaffold, then it doesn't generalize.”

Running RL at FP32 precision (especially at the LM head) is critical to closing the gap between theoretical and practical algorithm performance, a finding that cuts against common lower-precision training defaults.

“For example the precision part is one thing that we found that would you know kind of like prevent us from being close to that extreme.”

Olive Song on reward hacking during RL training at MiniMax.

“During reinforcement learning the model tries its best to hack a lot of things.”

MiniMax M2.7 and Kimi K3 signal a shift in open-weight licensing toward usage-based restrictions, moving away from permissive Apache-style terms.

“Now we're seeing a very healthy ecosystem development starting from even miniax recently when they're releasing the M2.7 model they have a term specifically focusing on usage and ki initially also has like if you have derivative works like this is kind of big news back then was fireworks and cursor about how they built on top of Kimi model.”

MiniMax ships a new model version roughly every 1 to 1.5 months, as of early 2026.

“I think we're launching one version about per month or a month and a half, right?”

Chinese AI vendors (Kimi, MiniMax, Tencent Cloud) immediately package open-source models as hosted commercial solutions, a playbook Kevin Xu notes US players are surprisingly not replicating.

“Which is why if you want to turn it to why Kimi and MiniMax and now Tencent Cloud and every other cloud under the sun in China is packaging Open Call almost immediately as a hosted solution to be consumed by the masses or whatnot.”

MiniMax runs an internal AI agent that monitors new papers, blogs, and articles, then automatically dispatches subject-tagged summaries and analyses to relevant researchers.

“We have a internal agent that tracks all the new articles and blogs and papers and then it dispatches to different subjects and then it summarizes and then it analyzes and researchers.”

MiniMax is forecast to release a trillion-plus parameter open model by end of 2026 that surprises observers on output quality, worth tracking against smaller prior releases.

“I think we will see a surprise from Miniax by end of the year. I think we will see a big model which like a really big model not M3 size but trillion parameters plus which will surprise us in terms of the outputs of miniax compared to before.”