OVHcloud open-sources Olala mini 7A1B hybrid-attention model
A 1B-active sparse model that runs on an 8GB GPU or plain CPU changes the calculus for teams deploying small models on their own infrastructure. Developers building local inference or edge deployments can compare it against other small open-weight models on throughput per euro of hardware. Its European training and Apache 2.0 licensing also give enterprises seeking non-US, non-China model provenance a new sourcing option.
IT之家Source