---
license: apache-2.0
base_model: openbmb/MiniCPM5-1B-Base
library_name: mlx
pipeline_tag: text-generation
language:
- en
- zh
tags:
- mlx
- minicpm
---

# MiniCPM5-1B-Base, 8-bit MLX

[openbmb/MiniCPM5-1B-Base](https://huggingface.co/openbmb/MiniCPM5-1B-Base) converted to MLX with
`mlx_lm.convert -q --q-bits 8 --q-group-size 64` (mlx-lm 0.31.3, affine). The weights are otherwise
unchanged, and the license is the upstream one (Apache-2.0).

Converted for [Pretype](https://github.com/nikiomori/Pretype), on-device autocomplete for macOS, where it
is the fastest completion model. On Pretype's 5,649-row, 17-language autocomplete set it ties the bf16
weights it replaces (first word right on 15% of all rows each; 14 against 21 discordant rows; English
28% of all rows for both), at 1.1 GB instead of 2.0 GB and about 30% less time per suggestion.

```python
from mlx_lm import load, generate
model, tokenizer = load("nikimoriOne/MiniCPM5-1B-Base-8bit")
print(generate(model, tokenizer, prompt="The quick brown fox", max_tokens=20))
```