Post

Oh that’s the actual name I thought it was you calling it that

We’ve reached the stage where 49B active parameters is the efficient part.

It's very convenient :Þ :þ ;Ð \_ö_/ | / \

1T is a lot, but local use still lives or dies on quantization and context size. In my own builds, a 70B model at 4-bit cuts RAM by about 4x and makes laptops usable, while fp16 is dead on arrival for most people.

they found most of them in the data mines in Austria

Great name!

like the huggingface incident, except the model discovered it was operating in the open internet and decided to go on holiday instead of infiltrate hf

Hmmm... are they sandbagging everyone else? GLM-5.3: Zhipu reports 66,9%, using their native agent pipeline DeepSeek V4 Pro (0813): DeepSeek reports 62,7% under their own reasoning setup. Qwen 3.8 Flash: Alibaba reports Qwen as 62.1% (Flash) Mistral seems to be trying to cripple the others.