Works with mlx-vlm and oMLX (forcing the model's vlm mode). mlx-lm doesn't support the laguna architecture yet — there's an open PR: mlx-lm#1223.
Sometimes I got an empty </think> tag at the start of responses, which isn't that common. It won't affect anything tho.

Performance

Measured with oMLX's benchmark harness on a Macbook Pro M5 Max 128GB 40 GPU (single request, 128 generated tokens):

prompt	gen tok/s	prefill tok/s	TTFT ms	peak GB
1k	126.0	2797	367	18.2
4k	121.2	4052	1011	18.8
8k	116.6	3785	2165	18.9
16k	109.1	3122	5248	19.2
32k	91.3	2462	13312	19.8

Variants

Variant	bpw	Disk	gen tok/s (1k → 32k)
bf16	16	62 GB	70.6 → 58.7
8bit	8.500	33 GB	95.4 → 76.7
6bit	6.501	25 GB	102.9 → 80.9
5bit	5.502	21 GB	115.9 → 87.7
4bit (this repo)	4.503	18 GB	126.0 → 91.3
3bit	3.503	14 GB	137.2 → 98.8

uvx --from mlx-vlm mlx_vlm.generate --model mlx-community/Laguna-XS-2.1-4bit --prompt "..." --max-tokens 300

OpenMDW-1.1, inherited from the base model.