[AkaiNa]

Yes. But standard block floating point uses a linear grid scaled by a shared exponent. Whereas AXS-6 uses a NormalFloat grid scaled by a shared exponent to maximize information density for bell-curve distributed weights. Essentially a Block Scaled Normalfloat-5.

[p1esk]

fp6 with block size 32 is a tough sell today when blackwell has native support for fp4 with block size 16.

How can I contact you?

[AkaiNa]

You can contact me on discord at brandon3183 or use the email registered with this account. Its less meant for data center scalability and more so meant for the everyday person who cant afford h100s and other data center scale gpu/npu since it supports any cuda gpu. Or for those who want to store larger models at home on a si gle or dual gpu configuration since eit uses less half the vram that bf16 does.