---
library_name: mlx
license: apache-2.0
base_model: IFM/K2-Horizon-MoVA-36B-A4B
tags:
- mlx
- apple-silicon
- mixture-of-experts
---

# K2-Horizon-MoVA-36B-A4B-MLX-4bit

4-bit affine (group size 64) MLX conversion of
[IFM/K2-Horizon-MoVA-36B-A4B](https://huggingface.co/IFM/K2-Horizon-MoVA-36B-A4B)
for Apple Silicon via `mlx-lm`.

The K2 MoVA architecture has no stock `mlx-lm` implementation, so this repo
ships a custom `k2_horizon.py`, referenced from `config.json` (`model_file`).
MoE/MoVA router matrices are kept in BF16; all large projections are 4-bit.

Measured on a 64 GB M4 Pro Mac mini: ~43 tok/s single-stream decode,
18.9 GB peak RSS, perplexity 2.1-8.6 on generic prose/code.

```python
from mlx_lm import load, stream_generate
model, tokenizer = load("iamEvanYT/K2-Horizon-MoVA-36B-A4B-MLX-4bit")
for chunk in stream_generate(model, tokenizer, "Hello.", max_tokens=64):
    print(chunk.text, end="", flush=True)
```