Quasar Ships Linear Attention, Plans SN3 Training

Quasar completed several technical milestones in April despite network delays, including Linear Attention outperforming Delta Attention and full architecture reveal using loop transformers. The team announced access to exclusive lab training data through a new collaboration and plans to train Quasar models on SN3 starting with a checkpoint on SN24. A decentralized MoE training run and distill incentive mechanisms are also in progress.

• •Linear Attention outperformed Kimi Delta Attention in benchmarks
• •Architecture fully revealed; uses loop transformers reportedly from Claude
• •Adaptation collaboration grants access to top-lab training data
• •SN3 multi-size model training collaboration underway