NanoForecast v0.5 vs TimesFM
6.5M-parameter edge model vs 200M-parameter foundation model. A 6.5M-parameter edge model beats a 200M-parameter foundation model on 4 of 6 standard benchmarks — after fixing 3 training pipeline bugs with zero architecture changes.
EXECUTIVE SUMMARY & EMPIRICAL VERDICT
NanoForecast v0.5 is 31× smaller, wins 4/6 benchmarks.
EULOGIK EDGE MODELFeatured
NanoForecast
v0.5 · 6.5 Million
Target HardwareApple M4 / CPU / Raspberry Pi
Latency / Speed19.5ms (CPU)
RAM / MemoryTiny
Fertility / SizeEdge Foundation Model
- ✓ 31× smaller than TimesFM yet wins 4 out of 6 benchmark datasets
- ✓ 43.8% MASE improvement (3.030 → 1.704) from training bug fixes
- ✓ CPU inference at 19.5ms on Apple M4
- ✓ Full reproducibility under Apache 2.0 (ArXiv 2609.31669)
INDUSTRY BASELINEOpen Weights
TimesFM
1.0 · 200 Million
Target HardwareGPU recommended
Latency / SpeedUnknown
RAM / MemoryUnknown
Fertility / SizeCloud Foundation Model
- • Leads on high-cardinality electricity and traffic datasets
METRIC SPECIFICATION MATRIX
| METRIC | NANOFORECAST | TIMESFM |
|---|---|---|
| ETTh1 MASE | 0.676 ★ | 0.705 |
| ETTh2 MASE | 1.110 ★ | 1.360 |
| ETTm1 MASE | 0.287 ★ | 0.545 |
| Exchange Rate MASE | 4.317 ★ | 4.383 |
| Electricity MASE | 3.512 | 2.180 ★ |
| Traffic MASE | 1.124 | 0.875 ★ |
| Overall MASE | 1.704 ★ | — |
REPRODUCIBLE CODE SNIPPET
Execute this benchmark locally
// See ArXiv 2609.31669 and GitHub: eulogik/NanoForecastInterested in deploying NanoForecast?
We assist enterprises with custom fine-tuning, hardware sizing, and on-premise air-gapped installation.