Benchmarks

Every benchmark, including the losses

Numbers with the artifact they were read from, and the gaps named.

Every benchmark on this site, including the ones we lose. Each links to the artifact the number was read from, and each says what was not measured.

Atome LM EDGE vs TensorFlow Lite for Microcontrollers

The most carefully run comparison here: UCI HAR, 2,947 held-out windows from 30 subjects, the same 6,567-parameter network through both engines, 5 seeds, leak-free subject-wise split.

Flash: WIN 4.64× Per-decision streaming: WIN 27.3× Accuracy: TIE Per-window latency: LOSS 2.89×

Read the full comparison

Language model: routed-ternary vs a plain float transformer

TinyStories, 3,000 steps, single seed. A direction rather than a result, and labelled that way.

RegimeAtome LMVanilla FP32Verdict
60K, parameter-fair6.31 ppl8.12 pplAtome −22%
60K, flash-fair6.31 ppl13.10 pplAtome −52%
944K, parameter-fair2.87 ppl2.54 pplAtome +11% — loses
944K, on disk271 KB3.7 MBAtome 20× smaller

Single seed, multi-seed run pending. The reversal at 944K is the honest headline: this architecture is a bet on the sub-1M regime and it loses outside it.

Verification, not comparison

CheckResult
Max |Δ|, Python → C99 → Cortex-M33.7e-7
Multi-token parity48/48 at 60K, 16/16 at 944K
C99 engine on all held-out EDGE windows0.9206, 99.19% agreement with float
Ternary-32 C engine vs reference2947/2947 identical

How parity is verified

Against the wider field

ProjectSmallestWeightsRuns on
BitNet b1.58700M – 3Bternaryserver / phone
llama2.c (Stories260K)260K – 110MFP32 / Q8MCU class possible
TinyMaix, esp32-llm~15MQ8 / FP32ESP32-S3 with PSRAM
Atome LM60K · 944Kternary, zero-heapCortex-M3 under QEMU, bit-exact; physical ESP32-WROOM-32

The claim is narrower than "best tiny LM" and more checkable: ternary weights, a zero-heap pure-C99 engine, and bit-exact parity verified under emulation. Each of those is falsifiable from the repository.

What is not benchmarked

The full list of gaps and losses