Qwen3.6-35B-A3B-TurboQuant-MLX-NVFP4

Summary

TurboQuant + MLX-NVFP4 (4-bit) variant of Qwen/Qwen3.6-35B-A3B.

Why this variant

Apple Silicon (M1/M2/M3/M4) with TurboQuant structural pre-conditioning and MLX-native NVFP4 layout (E2M1 weights, per-16-element FP8 (NVIDIA Blackwell layout)). 4.503 bits/weight, ~18 GB on disk, sub-2-s load on M4 Max. Pick this over the affine MLX variants when you want NVFP4 format parity with hardware pipelines while running locally.

Quickstart

import mlx.core as mx
from mlx_lm import load, generate

model, tokenizer = load("majentik/Qwen3.6-35B-A3B-TurboQuant-MLX-NVFP4")

prompt = "Give me a short introduction to Mixture-of-Experts models."
response = generate(model, tokenizer, prompt=prompt, max_tokens=512)
print(response)

[!IMPORTANT] Text-only extraction. This repo contains only the quantized language tower of Qwen3.6-35B-A3B. The upstream vision tower (333 tensors) and MTP head are not included, so image/video input does not work and mlx_vlm.load(...) fails with a Missing ... parameters error (the vision tower it expects is absent from the checkpoint). Load it with mlx_lm (recent version with qwen3_5_moe support) as shown above. For image/video input, use the upstream BF16 model Qwen/Qwen3.6-35B-A3B on a runtime that supports it.

Hardware compatibility

Device	VRAM	Recommendation
Apple M4 Max 128 GB	~21 GB	recommended — headroom for long context
Apple M3 Max 64 GB	~21 GB	fits comfortably
Apple M2 Max 32 GB	~21 GB	tight — short context only

Reproduce

# dequantize from the rotor/turbo MLX-8bit source, then re-quantize
python -c "from mlx_lm import convert; convert(hf_path=\"bf16\", mlx_path=\"out-nvfp4\", quantize=True, q_bits=4, q_group_size=16, q_mode=\"nvfp4\", trust_remote_code=True)"

Reproduced at commit 919836a.

Evaluation

Benchmark	Score
arc_easy_acc	0.4900
hellaswag_acc	0.4900

Family

bf16 — Qwen/Qwen3.6-35B-A3B
FP8 card — majentik/Qwen3.6-35B-A3B-FP8
TurboQuant MLX-4bit (affine) — majentik/Qwen3.6-35B-A3B-TurboQuant-MLX-4bit
plain MLX-NVFP4 (no rotor/turbo) — majentik/Qwen3.6-35B-A3B-MLX-NVFP4

Provenance

Calibration hash: none (nvfp4 is calibration-free; rotor/turbo conditioning inherited from source)
Uploaded: 2026-04-21T06:26:36.131649+00:00

Toolchain:

huggingface_hub: 1.11.0
mlx: 0.31.1
mlx-lm: 0.31.2

License

Released under apache-2.0. Upstream license of the base model applies.