T27 vs NVFP4: six weight formats on one small language model

Perplexity of six weight formats on SmolLM2-135M and the pre-registered ladder: T27 better than NVFP4, MX+ and MXFP4 on a trit substrate, worse than E2M2. Every limit listed.

This page sets no cookie and sends nothing anywhere.

Where the numbers come from

trinity-fpga research/block/ternary_storage_search_v2.json @ d31d8523cd (sha256 969e7b196f6afb9a...): the runner's output, perplexity per arm and split, permille ratios and verdicts; python 3.12.12, torch 2.11.0, transformers 4.57.6, run 2026-09-27 11:44:02Z to 11:52:11Z, not retried
trinity-fpga specs/numeric/ternary_storage_search_v2.t27 @ 49a84ade57 (sha256 88fcf517296ecc77..., the sha the run recorded): the pre-registration, compiled with the wasm t27 compiler for BEAT_PERMILLE, TIE_PERMILLE, NWIN, SEQLEN and the trits and bits per 32 weights
trinity-fpga research/block/TERNARY_STORAGE_V2_VERDICT_2026-09-27.md @ d31d8523cd: the report; its perplexity and ratio tables are cross-checked against the JSON, and every limit on this page is from its section What it does not say
node scripts/widget-data/t27-vs-nvfp4.mjs writes data.json; --check re-reads the three files and fails if data.json or a K_ fact here disagrees; card.png is drawn from data.json by scripts/widget-data/t27-vs-nvfp4-card.py

This widget's .t27 source · All widgets