Lila Sciences and Anyscale scale GLM-5.3-Flash to 1M tokens

Lila Sciences and Anyscale successfully scaled the training of the open GLM-5.3-Flash model to a one-million-token context length using 64 B200 GPUs. The collaboration utilized an open-source stack in SkyRL and implemented multiple memory and throughput optimizations, including kernel padding and distributed scoring for sparse attention, to handle the vast computational demands of long-context scientific data such as genomic sequences. Open article →