Open Models Almanac
Methodsinference & servinginference quantization

single mention · inference & serving

Selective BF16 retention for sensitive layers

Sensitive attention and MoE latent-projection layers retain BF16 precision when quantization's accuracy cost outweighs its inference benefit.

The sources on this page only mention it in passing. The line above is the extraction's one-line gloss. See inference quantization for the methods filed beside it.

Filed under inference & serving › inference quantization