WRLD

Aggregated public profile · Not an official account.

Octave Klaba Wrld summary

This profile is not managed by Octave Klaba.

Wrld compiles it automatically from public sources:

  • posts from X, summarised in the third person
  • quotes from the media
How it works ›
French entrepreneur, OVH owner and chief technical officer
Octave Klaba Wrld summary 4d

Octave Klaba announced the first model from OVH's LLM family, Olala mini 7A1B. He described it as a compact model with 7 billion parameters, with one billion being activated among 256 experts per layer, based on a hybrid attention architecture. According to him, the model could run on a GPU with at least 8 GB of memory, achieving 160 tokens per second on an RTX A4500 and 50 tokens per second on a processor. Its context window was 32,000 tokens. Klaba stated that Olala mini would be published on Hugging Face under the Apache 2.0 license. He specified that the model had been trained on the European public supercomputers Jupiter, MeluXina, and Jean Zay, and that this support represented the equivalent of 7 million euros in computing hours. According to Klaba, Qwant was already using Olala mini for its flash response feature, with responses 20% faster and costs reduced by three times compared to the previous model. He also mentioned that version 1.1, expected by the end of the year, was set to increase the context window to 128,000 tokens and improve mathematical reasoning. Version 2.0 would target enterprise usage, in cloud DevOps and cybersecurity, and exceed 100 billion parameters. https://t.co/3QlWMzfpkk #IA #LLM #OpenSource (translated)

3259