We've kept pushing our Open-R1 project, an open initiative to replicate and extend the techniques behind DeepSeek-R1

And even we were mind-blown by the results we got with this latest model we're releasing: ⚡️OlympicCoder

[1/3]

This phone made my night in the forest back in 2019

Thanks you to all my Qualcomm colleagues who helped with this research: Andrii Skliar, Ties van Rozendaal, Todor Boinovski, Mart van Baalen, Markus Nagel, Paul Whatmough, Babak Ehteshami Bejnordi

Proud to present our work on optimizing Mixture of Experts models for on-device generation speed: arxiv.org/pdf/2412.00099

We introduce a cache-aware routing that reduces cache misses of MoEs, improving token generation throughput by 2×. Perfect for memory-constrained devices.

I use pixi which is a conda env manager which can handle multiple envs (see « feature »). Each env can have a different set of pypi dependencies. ‘pixi —env cuda11.7 run …’ to run within the chosen env. Pixi uses uv under the hood to handle pypi dependencies.