Where does the bit budget go?

Qwen 3.8 27B
Measured10 bits / weight

Click a labeled budget for a measured allocation; drag between them to see the pattern interpolate. The budget is parameter-weighted: NVFP4 costs 4.5 bits with block scales, FP8 costs 8, and BF16 costs 16.

NVFP4FP8BF16Not present
Transformer layer →0481216202428323640444852566063
MLPGate
Up
Down
Linear attentionQKV
Z
Output
A
B
Conv1D
Full attentionQ
K
V
Output

Scroll across layers →

Layer 32 · MLP gateBF16Measured allocation