Efficiency and Optimization Breakthroughs
Key Questions
What efficiency improvements does Cohere Command A+ offer?
Cohere Command A+ is an open MoE model that achieves lossless W4A4 quantization. This reduces memory and compute requirements without sacrificing performance.
How do Sakana and NVIDIA achieve high sparsity on Hopper?
They demonstrate 99% sparsity techniques optimized for Hopper GPUs. This dramatically cuts compute needs while maintaining model quality.
What compute savings does OlmoEarth v1.1 provide?
OlmoEarth v1.1 delivers a 3x reduction in compute usage through targeted optimizations. Additional work on HCTA accelerators and spectral diffusion methods further addresses compute walls.
Cohere Command A+ open MoE with lossless W4A4; Sakana/NVIDIA 99% sparsity on Hopper; OlmoEarth v1.1 3x compute cut; HCTA accelerator and MLA/Spectral Diffusion target compute walls.
Sources (2)
Updated May 21, 2026