aadyaaadya-m1-mini
The tiny sibling. 0.84M parameters and 3.4 MB, built for phones, browsers and low-power devices, with the same API as aadya-m1.
At a glance
Small enough to ship inside an app.
0.84M
parameters
about 300 times smaller than aadya-m1
3.4 MB
weights in fp32
836,093 parameters
Same API
inputs and outputs
swap one repo name
Apache-2.0
licence
code and weights
Accuracy
Level with aadya-m1.
On the public real data we have, the two score the same. Choosing the small one costs nothing in accuracy there.
aadya-m1-mini against aadya-m1
Mean CRPS. Lower is better.
Real users: mini
1.932
Real users: aadya-m1
1.930
Real users: Bayes ref.
1.952
Simulated: mini
2.906
Simulated: aadya-m1
2.925
Simulated: Bayes ref.
3.026
View the numbers
| Method | Value | 95% interval |
|---|---|---|
| Real users: mini | 1.932 | - |
| Real users: aadya-m1 | 1.930 | - |
| Real users: Bayes ref. | 1.952 | - |
| Simulated: mini | 2.906 | - |
| Simulated: aadya-m1 | 2.925 | - |
| Simulated: Bayes ref. | 3.026 | - |
Real users: the two differ by less than the noise. Simulated users: the mini scores slightly lower, but each size was trained once, so we do not read that as a real advantage. Against the Bayes reference the mini is slightly ahead on real users.
Use it
Same code, different repository.
Choose the mini when
- The model ships inside a phone app, a browser or a low-power device
- Download size matters more than headroom
- You want the same forecasts without a heavy runtime
Choose aadya-m1 when you want the extra capacity to fine-tune on much more real data.
# Python 3.10+. Weights are downloaded once, then everything runs offline.
pip install numpy scipy torch safetensors huggingface_hub