Brooooooklyn/Qwen3.8-27B-NVFP4-mlx

Qwen3.8-27B — Unsloth NVFP4 + E4M3 FP8 weights (MLX, DGX)

Brooooooklyn/Qwen3.8-27B-NVFP4-mlx is an MLX mixed-weight-format quantization of Qwen/Qwen3.8-27B, prepared for experimental NVIDIA CUDA inference on Linux aarch64. The model has a 64-layer dense Qwen3.5-family language backbone with 48 linear-attention and 16 full-attention layers, a BF16 vision tower, and one preserved MTP layer.

This model is part of the Unsloth NVFP4 Tensor-Class Recipe for MLX — macOS + DGX collection.

The source was the all-BF16 checkpoint at revision 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0.

Update — re-converted 2026-09-01

Every dense SwiGLU FFN block scale is lifted by an exact power of two, folded into the feeding norm. NVFP4 stores a block scale as amax / 6 in E4M3, and on real weights those land in E4M3's subnormal band, carrying three mantissa bits instead of four -- or rounding to the zero code, which decodes all 16 weights in the block to zero. The fold is exact, so the layer's function is unchanged.

Annihilated dense MLP blocks: 0.0049% -> 0.0000%; worst dense tensor 0.0770% -> 0.0000%.

Weight bytes changed; the tensor-class map, shard layout and tensor counts are unchanged. Earlier revisions remain in this repository's history.

Quantization recipe

This is a data-free, weight-only MLX storage port of the Unsloth Qwen3.6 NVFP4 recipe, applied to Qwen3.8's matching dense Qwen3.5-family tensor layout:

• tensors assigned NVFP4 by the recipe are stored as NVFP4, 4-bit with group size 16;
• tensors assigned FP8 by the recipe are stored as raw E4M3 FP8 weight bytes with one BF16 dequantization scale per output channel;
• every excluded tensor stays BF16.

The NVFP4 class uses MLX weight-only quantized matmul with BF16/A16 activations. For the FP8 class, mlx-node reconstructs each weight to BF16 once at load and then uses ordinary A16 matmul. The serialized fp8_e4m3 form is Uint8 [N, K] weight plus BF16 [N, 1] scale; it is not MLX mxfp8 and it is not native W8A8 execution.

No imatrix, calibration dataset, AWQ-style pre-scaling, activation calibration, NVFP4 global scale, or FP8 KV-cache calibration was used. This artifact preserves the recipe's tensor-class boundaries and weight storage formats under mlx-node's A16 runtime; it does not claim numerical or performance parity with Unsloth's calibrated W4A4/W8A8 execution.

Tensor class	Stored format
Dense FFN {gate,up,down}_proj, layers 0–55	NVFP4 4/16
Dense FFN {gate,up,down}_proj, layers 56–63	E4M3 FP8 weight + per-output BF16 scale
Full-attention {q,k,v,o}_proj	E4M3 FP8 weight + per-output BF16 scale
Linear-attention in_proj_qkv, in_proj_z, out_proj	E4M3 FP8 weight + per-output BF16 scale
lm_head	E4M3 FP8 weight + per-output BF16 scale
Embeddings; in_proj_a/b; GDN state, convolution, and norm tensors; all other norms	BF16
Entire 15-tensor mtp.* subtree	BF16
Vision tower and merger tensors	BF16

The allocation contains 168 NVFP4 modules and 233 E4M3 FP8 modules. The top-level config is nvfp4, 4-bit, group size 16; the 233 higher-class modules are explicit fp8_e4m3 overrides. The final eight dense FFNs intentionally use the higher class.

Target and usage

The canonical native target is aarch64-unknown-linux-gnu: Linux aarch64 with glibc and NVIDIA CUDA 13.0. mlx-node currently validates this experimental inference-only path on NVIDIA GB10 / DGX Spark (sm_121). It is not a generic CUDA or x86_64 artifact, and there is no published CUDA prebuilt native npm binary. Build the pinned checkout on the DGX host:

git clone https://github.com/mlx-node/mlx-node.git
cd mlx-node
git checkout e281f0bbc61eb71a244ad6681ffad324a1177780
git submodule update --init --recursive
yarn install
yarn build

The current CUDA preview uses eager execution and a flat KV cache:

MLX_QWEN35_FORCE_EAGER=1 \
MLX_QWEN35_PAGED_OVERRIDE=0 \
  yarn oxnode your-script.ts

For example:

import { loadSession } from '@mlx-node/lm';

const session = await loadSession('./Qwen3.8-27B-NVFP4-mlx');
const result = await session.send('Explain the purpose of a unit test in one sentence.');
console.log(result.text);

The full BF16 MTP subtree is preserved inline and is supported by mlx-node on the macOS fallback path. CUDA speculative decoding is not supported in the DGX preview, so preservation of these tensors does not claim DGX MTP validation.

Reproduction

Converter checkout: mlx-node e281f0bb (package version 0.0.10). The invocation was:

mlx convert \
  --input /Users/brooklyn/.mlx-node/models/qwen3.8-27b \
  --output /Users/brooklyn/.mlx-node/models/qwen3.8-27b-unsloth-nvfp4-mlx \
  --model-type qwen3_5 \
  --dtype bfloat16 \
  --quantize \
  --q-recipe unsloth \
  --q-mode nvfp4

No --imatrix-path was supplied. The converter therefore applied the fixed tensor-class map without AWQ pre-scaling.

Validation

The five-shard SafeTensors index contains 1,600 entries and reports metadata.total_size = 23,417,338,336 bytes. Header and index validation confirmed exact shard closure, valid physical offsets, 168 inherited NVFP4 groups, 233 complete fp8_e4m3 groups, 401 scale sidecars, no bias sidecars, and identical quantization and quantization_config blocks.

All 333 vision/merger tensors and all 15 mtp.* tensors remain BF16 without quantization sidecars. No imatrix or calibration artifact is present. The tokenizer, chat template, generation config, and preprocessor assets are byte-identical to the pinned source snapshot.

A macOS A16-fallback smoke test detected hasMtpWeights() = true and completed a short deterministic generation with native MTP explicitly enabled. This plumbing check does not validate model quality, long-context behavior, vision quality, DGX/CUDA execution, or benchmark performance.

License and attribution

The source model card declares the Apache-2.0 license. Model capability and training credit belong to the Qwen Team. The tensor-class recipe is credited to Unsloth. This repository converts the pinned BF16 source weights into the mixed NVFP4/plain-E4M3 MLX representation described above.