Introducing Olmo-core 3: Open, scalable training infrastructure for large MoEs

In 30 Sekunden

Was passiert ist

Allen AI released Olmo-core 3, an open training framework for mixture-of-experts models scaling to trillion parameters while maintaining computational efficiency.

Warum es zählt

Matters for researchers and engineers building large sparse models who need efficient distributed training infrastructure across GPU clusters.

Achtung

Benchmarks used random routing rather than actual trained models; sustained performance at trillion-parameter scale remains to be demonstrated in full training runs.