mike

[mike @mickyyy68]

@Mirjalolfl @TencentHunyuan It is 770B and its weight in BF16 are 1.5TB

[mike @mickyyy68]

@abue_ammar will try it asap

[mike @mickyyy68]

@abue_ammar straight fire

[mike @mickyyy68]

@EggMasonValue @dhh @tobi @trq212 @agents fuck anthropic

[mike @mickyyy68]

@EggMasonValue @dhh @tobi @trq212 @agents or you could just symlink it

[mike @mickyyy68]

@Yuchenj_UW K3 isn't cheap and is also really good at writing kernels

(source: their blog for K3)

[mike @mickyyy68]

how is grok 4.6 high performing better than xhigh lmao

[mike @mickyyy68]

@cabralorenzo tanta roba

[mike @mickyyy68]

@burkeholland Grok. Really cool UI and UX. The design is so good and the speed/research is sota

[mike @mickyyy68]

@abue_ammar yeah maybe it could be time to bring in the markdown formatter xd

[mike @mickyyy68]

@abue_ammar what would u like it to be?

[mike @mickyyy68]

@TechProphet_X @ggerganov yeah but it isn't close

compared to the 35b a3b you are 27b vs 3b for each token, so you can easily see the difference in quality

like also with the 3.6, the 27B was better than the 35b (and it isn't that close tbh)

[mike @mickyyy68]

@OrganicGPT @ggerganov your experience is only benchmarks? in the graph it beats deepseek

[mike @mickyyy68]

@TechProphet_X @ggerganov I do not think that a moe could get the performance of this dense one

[mike @mickyyy68]

@oxo_nomy @nahcrof up to how much? (just curious, not hating)

[mike @mickyyy68]

@nahcrof I really appreciate your mission in today society

though I think 75% is still low, to feel the "cheapest" inference in the world it should be at least 95%

I am not hating, I really appreciate what you are doing, just giving my pov

[mike @mickyyy68]

@stevibe what about cache hit?