eulogik/nanoforecast-200k

🔮 NanoForecast 200k (v0.1)
_{World's smallest deployable time series transformer}

676K params · 256 context · Just 2.7 MB
Fits on a Raspberry Pi Zero. Ships as a dependency, not an API call.

💡 Need better accuracy? Check nanoforecast-500k (1.6M params, MASE 3.45) or the latest nanoforecast-v03 (6.5M params, MASE 2.73).

---

📦 Quick Start

pip install nanoforecast
python3 train_from_csv.py --csv my_data.csv --target sales --horizon 48

Or use a pretrained model:

from nanoforecast import NanoForecast
model = NanoForecast.from_pretrained("eulogik/nanoforecast-200k")

📋 When to Use This Model

Scenario	Pick This Model
Raspberry Pi Zero (512 MB RAM)	✅ nanoforecast-200k
Browser via ONNX.js (< 1 MB)	✅ nanoforecast-200k
AWS Lambda cold starts	✅ nanoforecast-200k
Battery-powered IoT sensor	✅ nanoforecast-200k
Any accuracy-critical app	❌ Use v0.3 instead

🔄 Streaming Inference

Feed one value at a time with DeltaNet stateful RNN:

result = model.predict(context, horizon=48, return_state=True)
state = result.pop("state")
for new_val in incoming_data_stream:
    result = model.predict_step(new_val, state, horizon=48)

📚 All Variants

Model	Params	Context	Size	Best For
nanoforecast-200k (you are here)	676K	256	2.7 MB	Extreme edge
nanoforecast-500k	1.6M	256	6.4 MB	General purpose
nanoforecast-v03	6.5M	512	26 MB	Max accuracy

❤️ Built by Eulogik

Eulogik — deployable AI for the real world.

Star the repo ⭐ on GitHub!