kronos-mlx-base / README.md
gxcsoccer's picture
Initial MLX-native Kronos-base weights
52f80e9 verified
---
license: mit
library_name: kronos-mlx
pipeline_tag: time-series-forecasting
tags:
- mlx
- apple-silicon
- finance
- kronos
- time-series
base_model: NeoQuasar/Kronos-base
---
# Kronos-base (MLX)
Apple [MLX](https://github.com/ml-explore/mlx) port of [`NeoQuasar/Kronos-base`](https://huggingface.co/NeoQuasar/Kronos-base) — the 102M-parameter Kronos variant. d_model=832, n_layers=12, max_context=512. Pair with [`gxcsoccer/kronos-mlx-tokenizer-base`](https://huggingface.co/gxcsoccer/kronos-mlx-tokenizer-base).
## Usage
```python
from kronos_mlx import Kronos, KronosTokenizer, KronosPredictor
tokenizer = KronosTokenizer.from_pretrained("gxcsoccer/kronos-mlx-tokenizer-base")
model = Kronos.from_pretrained("gxcsoccer/kronos-mlx-base")
predictor = KronosPredictor(model, tokenizer, max_context=512)
```
For 8-bit Linear weight quantization (390 MB → ~115 MB, **-71 %**):
```python
model = Kronos.from_pretrained("gxcsoccer/kronos-mlx-base", bits=8)
```
8-bit on Kronos-base is much higher fidelity than on Kronos-small thanks to the larger model's redundancy — recommended for memory-constrained Apple Silicon.
## Original
- Upstream: [shiyu-coder/Kronos](https://github.com/shiyu-coder/Kronos) (AAAI 2026)
- PyTorch weights: [NeoQuasar/Kronos-base](https://huggingface.co/NeoQuasar/Kronos-base)
- MLX port: [github.com/gxcsoccer/kronos-mlx](https://github.com/gxcsoccer/kronos-mlx)