Summary
The result compares with 11.1 times for GLM-5.2 and reportedly passes the benchmark’s single-launch gate that the earlier version failed. The implementation used Kimi-Linear Decode on an RTX PRO 6000 and CUDA techniques including persistent CTAs, Int4 dequantisation and fused latent-attention handling.
Why it matters
It points to inference optimisation becoming a model-capability frontier, with software and kernel design materially changing usable performance.
Related coverage
- SGLang and AntLing cut trillion-parameter restart time to 32 seconds
- Token fingerprints link anonymous Ox Alpha to Baseten’s GLM-5.2 Vision
- Microsoft receives first production Vera Rubin systems as Nvidia ramps the platform