Sub-1-bit quantization crosses from theory to method — Samsung NanoQuant
Samsung's NanoQuant (arXiv 2602.06694, ICML 2026) is the first post-training quantization to reach both binary and sub-1-bit compression with calibration only (128 samples, 1 GPU, no retraining): a 70B model drops from 138 GB to 5.35 GB and runs on a consumer 8 GB GPU at ~20 tok/s. Below-one-effective-bit-per-weight is now a real technique, not a whiteboard idea — the direction Alex Wissner-Gross's 'sub-1-bit goes mainstream' extrapolation points at.
Why it matters (spine — atoms). Doing more with fewer bits, out to the edge. Assume the cost/size floor of on-device intelligence keeps falling ~10× on a 12-18 month horizon; design edge products against next year's floor, not today's.
What would prove it wrong (falsifier). Sub-1-bit models fail to hold accuracy on real downstream tasks, leaving ~1.58-bit ternary as the practical floor.
Live-status check + triangulation. Samsung's NanoQuant (arXiv 2602.06694, ICML 2026) is the first post-training quantization to reach both binary and sub-1-bit compression with calibration only (128 samples, 1 GPU, no retraining): a 70B model drops from 138 GB to 5.35 GB and runs on a consumer 8 GB GPU at ~20 tok/s. Below-one-effective-bit-per-weight is now a real technique, not a whiteboard idea — the direction Alex Wissner-Gross's 'sub-1-bit goes mainstream' extrapolation points at. Source: Paper · Samsung / ICML 2026 (2026-02-01); see source link. Every post-cutoff figure was tool-fetched against a live source during the Kimi K3 sweep (2026-07-20); see sweeps/2026-07-20-VERIFIED_DIGEST.md for the full provenance and any corrections of record.
Assume the cost/size floor of on-device intelligence keeps falling ~10× on a 12-18 month horizon; design edge products against next year's floor, not today's.
domains: edge-compute · quantization · atoms