Nanbeige4.3-33B-MLX-4bit

A 4-bit MLX conversion of Nanbeige/Nanbeige4.3-33B, converted with mlx-lm for Apple Silicon.

All credit for the model itself goes to the Nanbeige Team. This repo is only a format conversion of their Apache-2.0 release — no weights were trained, tuned, or modified beyond quantization.

Details

• Base model: Nanbeige/Nanbeige4.3-33B (33B dense, Llama architecture, 512K context)
• Quantization: MLX affine, 4 bits (4.5 bpw effective), group size 64
• Size on disk: ~18.6 GB (4 shards)
• License: Apache-2.0 (same as the base model)

Why this model

Nanbeige4.3 is a thinking model tuned for agentic coding work. Vendor-reported scores from the base model card (same harness, vendor-run):

Benchmark	Nanbeige4.3-33B	Qwen3.8-27B
Terminal-Bench 2.1	80.9	73.0
DeepSWE	65.5	42.5
Automation Bench	54.2	38.3
SWE-Bench Pro	60.3	61.7
GPQA Diamond	89.5	—

The base card recommends tool_call_format="xml" and, for agent work, keeping thinking enabled with preserve_thinking on.

Use

LM Studio: download this repo into your LM Studio models folder (or use lms get Duckets/Nanbeige4.3-33B-MLX-4bit), then load it like any other local MLX model.

mlx-lm:

pip install mlx-lm
python -m mlx_lm.generate --model Duckets/Nanbeige4.3-33B-MLX-4bit --prompt "Write a Python function that retries a request with exponential backoff."

Verification

• Converted from the original 14 safetensors shards with mlx_lm.convert -q.
• Smoke-tested with mlx-lm: coherent thinking trace, correct answer on a numeric-comparison probe (~30 tok/s generation, 18.8 GB peak memory on an M-series Mac).
• LM Studio 0.4.26: indexed and loaded (nanbeige4.3-33b-mlx), generation-tested on a Mac Studio — it works. Honest performance note: this is a dense 33B with no Splash/DFlash-style acceleration, so on Apple Silicon it generates noticeably slower than MoE models of similar size. That's the hardware/format reality, not a defect in the conversion.