Four-bit mix shrinks Cellpose-SAM weight storage 6.8x with no field failures

single source· 1 articles · confidence: high · first seen 2026-09-16 20:00 UTC

What this means for you

If you run microscopy segmentation on lab CPUs, the transferable part is the method: fix a tolerance before compressing and check it per imaging modality. The 4-bit mix is a candidate to test, not a default — this is one preprint using its own criterion, and ternary weights are out.

A preprint tests how far Cellpose-SAM, a segmentation model (it labels individual cells in microscope images) used in stem-cell microscopy, can be compressed to run on laboratory CPUs. Quantisation stores weights at lower numeric precision to cut memory. On a 176-field panel across three imaging sets, the authors require in advance that the 95% cluster-bootstrap interval of mean change from the FP32 baseline stay above -0.02 in every modality. 8-bit weights with 16-bit activations hold instance F1 throughout. A 4-bit mix with four INT8 exceptions cuts weight storage 6.76x with no failures in 176 fields; ternary weights cut 12.08x but fail on 169.

Key facts

  • ·Paper is arXiv 2609.21038, posted 16 September 2026 source
  • ·Evaluation panel is 176 fields spanning BBBC038 nuclei, BBBC039 U2OS fluorescence and NIST iPSC images source
  • ·Pass criterion: the 95% cluster-bootstrap interval of mean change from FP32 must stay above a -0.02 margin for every imaging modality source
  • ·Weight-only W8A16 preserves instance F1 across all modalities; a mixed W4/W8 scheme using four INT8 exceptions gives a 6.76x reduction in weight storage with 0 catastrophic failures in 176 fields source
  • ·Ternary weight-only quantisation gives 12.08x compression but fails catastrophically on 169 of 176 fields source

What the sources say

Sources

The original reporting. Follow these — they did the work.

← the wire