MLX NVFP4 (nvfp4, 4-bit, group size 16) quantization of Qwen/Qwen3.6-27B at upstream revision 6a9e13bd6fc8f0983b9b99948120bc37f49c13e9, produced directly from the BF16 safetensors with pipelines.mlx_direct_quantize (majek repo). ~15.2 GiB on disk.

Runtime caveat: this is a weights + config pack published ahead of runtime availability; the qwen3_5 vision-language architecture is now supported upstream (mlx-lm PR #1345, merged 2026-06-04; mlx-vlm support available). No local inference smoke was performed on these artifacts, so the quality of this variant is not verified.

Why this variant

MLX-native NVFP4 — E2M1 weights with per-16-element FP8 scales (NVIDIA Blackwell layout), group size 16.

Reproduce

python -m pipelines.mlx_direct_quantize --model qwen3.6-27b --base-dir /tmp/mlx-direct-release/qwen3.6-27b/base --out-dir /tmp/mlx-direct-release/qwen3.6-27b/NVFP4 --bits 4 --mode nvfp4 --group-size 16

Vision tower

The model.visual.* tower (333 tensors) is passed through unquantized in BF16 — only the text tower (model.language_model.*, lm_head.*) is quantized.

Family

All MLX variants of this model ship together:

majentik/Qwen3.6-27B-MLX-2bit
majentik/Qwen3.6-27B-MLX-3bit
majentik/Qwen3.6-27B-MLX-4bit
majentik/Qwen3.6-27B-MLX-5bit
majentik/Qwen3.6-27B-MLX-6bit
majentik/Qwen3.6-27B-MLX-8bit
majentik/Qwen3.6-27B-MLX-MXFP4
majentik/Qwen3.6-27B-MLX-NVFP4 (this repo)

Provenance

Upstream: Qwen/Qwen3.6-27B @ 6a9e13bd6fc8f0983b9b99948120bc37f49c13e9
Quantization: bits=4, mode=nvfp4, group_size=16 (text tower only)
Toolchain: mlx 0.31.2, huggingface_hub 0.36.2
Full details in PROVENANCE.md in this repo.

License + attribution

Quantized by majentik from Qwen/Qwen3.6-27B. All rights in the original model remain with its authors.

The upstream model is released under the Apache License 2.0 — see the upstream LICENSE file.

majentik/Qwen3.6-27B-MLX-NVFP4

Qwen3.6-27B-MLX-NVFP4

Summary