X · @teortaxesTex
· X / Twitter
RT xjdr: we've been running GLM 5.2 in bf16 and in fp8 (experts and kvcache only, attention is always bf16) and have recorded virtually 0 measurable q…
RT xjdrwe've been running GLM 5.2 in bf16 and in fp8 (experts and kvcache only, attention is always bf16) and have recorded virtually 0 measurable quality difference in our A/B tests and audits (very surprisingly)NVFP4 has shown a slight performance regression but could probably be fixed with proper L/DoRA and might al