3 ms·
Why not? The model is 2.8T parameters with native MXFP4, which is 1400GB.
by johndough 3mo ago
Why not? The model is 2.8T parameters with native MXFP4, which is 1400GB.
- alightsoul 3mo agoI would figure it needs a lot of kV cache and context size which can be much bigger than the model itself