Lossless BF16-to-12-bit model compression ships with efficient NyoomFloat12 kernels
MaleManlpulator · x · 2026-11-07
AlpinDale unveiled a lossless compression method that shrinks BF16 models to 12 bits, built on Together AI's NyoomFloat12. Since no hardware natively supports this dtype, significant engineering went into making it run efficiently. A commenter asks why the approach can't extend further, e.g. to FP32.
Related event: Developer open-sources nf12: lossless 12-bit compression for BF16 models(10 posts)→