cgs4huggingface/

Qwen3.6-27B-MLX-VL-oQ5-fp16

❤️ Thanks to deepsweet for the fantastic work on this readme and all uploads of quantized models.

This model was converted to MLX format and quantized from Qwen3.6-27B using oMLX.

See "oQ: oMLX Universal Dynamic Quantization" for details.

Unlike the Qwen3.6-35B-A3B collection, work on the full Qwen3.6-27B 4x4 table is still ongoing. So if you see a 404 link below, feel free to ask deepsweet to prioritize it.

"VL" is Vision-Language, meaning quantization preserves the original model's multimodality.

No "VL" means quantization is Text-Only.

"FP16" is an M1/M2 Apple Silicon tweak that delivers a very noticeable prompt processing boost, because older M-series lack native BF16 hardware support. See "Metal FP32 Vs BF16 Vs FP16 benchmark" for details.

No "FP16" means quantization is better suited for M3+ Apple Silicon.