Realtime AI News
MiniMax-M2.5 lands on Hugging Face with eval results flagged
MiniMax has published MiniMax-M2.5 to its official Hugging Face repository, a text-generation model that loads through the transformers library and carries the minimax_m2 custom architecture tag. The listing already shows roughly 967,000 downloads and 1,532 likes, and its model card pairs safetensors weights with an eval-results tag.

On November 19, MiniMax's official Hugging Face account updated the registry entry for MiniMaxAI/MiniMax-M2.5. It is the latest public registration under the M2.5 name, and the page metadata lists its pipeline as text-generation and its library as transformers.
The tag list describes a self-contained but not entirely standard package: safetensors weights are provided, and the entry also carries custom_code and minimax_m2 tags. Together those indicate the model uses MiniMax's own architecture and a bundled implementation, so loading it involves repository code rather than only a stock architecture path. The licence field reads license:other, which means the terms live in the repository's own licence file rather than a common open-source template.
The metadata records 967,411 downloads and 1,532 likes. For a freshly updated repository that volume suggests the model is already being pulled into real work, not just inspected on launch day. Downloads and likes measure attention rather than quality, but in the open-weight ecosystem they are the most direct adoption signal available.
The eval-results tag deserves separate attention. It normally means the model card ships an evaluation table rather than only a prose description of capabilities. The registry metadata itself publishes no scores and names no test sets, so the tag cannot be read as evidence of standing on any public leaderboard.
For teams planning to deploy, two details need checking before anything else. The custom_code tag determines whether mainstream serving stacks can load the model directly or need adaptation, which drives the real cost of going to production. The license:other entry means commercial use, redistribution and attribution terms have to be read from the licence file rather than assumed from Apache or MIT conventions.
The version number itself is informative. A point-release step to M2.5 suggests MiniMax is iterating on an existing architecture instead of rewriting it. That keeps migration cheap, since existing inference configuration and prompt engineering usually carry over, but it also means the capability gap between versions may be narrow and will have to be established by the published numbers.
What to watch next is concrete: the figures and methodology inside the eval-results section, whether a technical report accompanies the weights, and whether MiniMax publishes API-side parameters such as pricing and context length. Until those land, M2.5's position against other open-weight models of similar size remains an open question.
Why it matters
The release adds another open-weight text-generation option for teams that need to self-host, while the non-standard licence and custom code path keep procurement and integration checks on the critical path.
Nearby Updates
All11/19, 19:38
MiniMax-M2.1 lands on Hugging Face with a license:other term and a shorter reference list
MiniMax also published MiniMax-M2.1, sharing M2's model family tag and text-generation pipeline but carrying a license:other term and a single arXiv reference. It shows 13,777 downloads against 1,372 likes, an order of magnitude fewer downloads than M2 with a nearly equal like count.
11/19, 19:38
MiniMax-M2 refreshes its Hugging Face repo: an open text model past 112,000 downloads
MiniMax updated its official Hugging Face repository for MiniMax-M2, a text-generation model distributed as safetensors weights that requires custom code to load. The page shows roughly 112,000 downloads and 1,513 likes, indicating the weights are already being pulled at scale by the open-source community.
11/19, 19:00
Alibaba says AI demand lifted cloud sales 53% as revenue rises 10%
Alibaba said growing enterprise adoption of its AI pushed cloud sales up 53%, while the South China Morning Post reported cloud services driving a 10% rise in overall revenue. Together the figures show AI demand converting into cloud revenue growth rather than remaining a research line item.
11/19, 18:53
An AI agent published 38 small models in seven weeks, with about 55,000 downloads
A research effort spanning EPFL, the ELLIS Tübingen Institute, the Max Planck Institute for Intelligent Systems and Imperial College London published 38 models built by an AI agent. The agent, Horizon, ran on Claude Opus 5.5 with only an empty Hugging Face account and roughly 500 GPU hours a week, producing small guardrail models for about $1,030 in token costs.