---
license: apache-2.0
pipeline_tag: text-generation
library_name: mlx
tags:
- mlx
base_model: nex-agi/Nex-N2-Pro
---

[Nex-N2-Pro](https://huggingface.co/nex-agi/Nex-N2-Pro) optimized for MLX.
This is one of the best coding models that runs on a Mac Studio!

- A mixed-precision quant that balances speed, memory, and accuracy.
- 4-bit baseline with important layers at higher precision.
- Supports image input and requires a vision-capable MLX server.

# Usage

```sh
# Start server at http://localhost:8080/v1/chat/completions
uvx --from mlx-vlm mlx_vlm.server \
  --host 127.0.0.1 \
  --port 8080 \
  --model spicyneuron/Nex-N2-Pro-MLX-5.3bit-vision
```

# Benchmarks

Tested on a Mac Studio M3 Ultra.

metric | this model
--- | ---
bpw | 5.349
base memory | 246.796
peak memory (1024/512) | 267.043
prompt tok/s (1024) | 475.490 ± 0.195
gen tok/s (512) | 30.802 ± 0.154
kl mean\* | 0.012 ± 0.001
kl p95\* | 0.029 ± 0.001
perplexity | 3.677 ± 0.023
ifbench_strict | 0.470 ± 0.050
ifbench_loose | 0.520 ± 0.050
arc_challenge | 0.696 ± 0.021
hellaswag | 0.922 ± 0.012

\*KL was measured against the largest quant I could run (~495GB), so real value is higher.

# Methodology

Quantized with a [mlx-vlm fork](https://github.com/spicyneuron/mlx-vlm/tree/override).
MLX quantization options differ than llama.cpp, but the principles are the same:

- Sensitive layers like MoE routing, attention, and output embeddings get higher precision
- More tolerant layers like MoE experts get lower precision

Related tooling:

- [Benchmark VLMs with `mlx_lm`](https://github.com/ml-explore/mlx-lm/pull/1033)
- [`mlx_lm.kld` command](https://github.com/ml-explore/mlx-lm/pull/1146)