Downloads · 30 days
21
7% of all-time downloads
davisliu/Qwen3.6-27B-MXFP4-VL-MTP
Qwen3.6-27B-MXFP4-VL-MTP is a image-text-to-text model from davisliu. Use it for the image-text-to-text task on the model card, and read the license before you ship it in a product. It is set up for mlx. The card lists the license as apache-2.0.
Downloads · 30 days
21
7% of all-time downloads
All-time downloads
320
Public
Parameters
27.8B
16.1 GB on disk
Likes
1
Public
Click a slice to open those files.
.safetensors16.1 GB · 100%
How the weights are stored.
U3226.9B · 97%
From the Hugging Face model README
English | 中文
A vision-preserving MXFP4 quantization of Qwen3.6-27B for Apple Silicon, with the MTP (Multi-Token Prediction) head kept intact for native speculative decoding — no separate draft model required.
面向 Apple Silicon 的 保留视觉能力的 MXFP4 量化版本,完整保留 MTP(多 token 预测)头,原生支持投机解码——无需独立的草稿模型。
This is an MXFP4 (4-bit) MLX quantization of Qwen/Qwen3.6-27B, produced with junainfinity/mlx-mtp. Unlike generic 4-bit conversions, this build:
a_log, dt_bias, conv1d, ssm_*) to avoid hybrid-state degradationt+2 from the hidden state at position t| Property | Value |
|---|---|
| Base model | Qwen/Qwen3.6-27B |
| Quantization | MXFP4 (4-bit, group_size=32, packed uint32 + E8M0 scales) |
| Vision tower / projector | fp16 (not quantized) |
| MTP layers | 1 (quantization_config.mtp_num_hidden_layers: 1) |
| Parameters | ~27B |
| Context length | 262,144 |
| Architecture | Qwen3.5 (Gated DeltaNet hybrid), VL variant |
| File size | ~16GB (4 safetensors shards) |
| Toolchain | junainfinity/mlx-mtp, --mode mxfp4 --verify |
| Component | Precision | Notes |
|---|---|---|
| Language backbone | MXFP4 (4-bit, group_size=32) | packed uint32 weights + uint8 E8M0 scales |
| Vision tower / projector / merger | fp16 | full image capability preserved |
MTP head (mtp.*) | kept (fp16) | self-draft speculative decoding |
| SSM-sensitive params | high precision | a_log, dt_bias, conv1d, ssm_* |
Qwen3.6-27B-MXFP4-VL-MTP/
├── config.json # mlx_mtp metadata + quantization_config.mtp_num_hidden_layers: 1
├── generation_config.json
├── preprocessor_config.json # image processor config
├── chat_template.jinja
├── tokenizer.json
├── tokenizer_config.json
├── vocab.json
├── merges.txt
├── model.safetensors.index.json
├── model-00001-of-00004.safetensors # ~5.3GB
├── model-00002-of-00004.safetensors # ~5.3GB
├── model-00003-of-00004.safetensors # ~5.4GB
└── model-00004-of-00004.safetensors # ~0.2GB
Note: whether MTP is active is controlled by
quantization_config.mtp_num_hidden_layersinconfig.json(1= enabled,0/ missing = disabled).
Use the inference engine and OpenAI-compatible server from junainfinity/mlx-mtp (mlx-mtp-serve).
Load via VLMModelFactory in mlx-swift-lm with Qwen3.5/3.6 VL support — point it at this directory / repo directly.
mlx-mtp-quantize \
--src Qwen/Qwen3.6-27B-bf16 \
--out Qwen3.6-27B-MXFP4-VL-MTP \
--mode mxfp4 \
--verify
BF16 source ~55GB → MXFP4 ~16GB. Vision tower, MTP head and SSM-sensitive tensors are excluded from quantization by the tool's skip predicates.
mtp_num_hidden_layers > 0 (e.g. mlx-mtp, or mlx-swift-lm builds with Qwen3.5 MTP support)Apache-2.0, same as the base model. Provided as-is, without warranty.
这是对 Qwen/Qwen3.6-27B 的 MXFP4(4-bit)MLX 量化版本,使用 junainfinity/mlx-mtp 制作。与普通 4-bit 转换不同,本版本:
a_log、dt_bias、conv1d、ssm_*),避免混合状态退化t 的隐藏状态预测 token t+2| 属性 | 值 |
|---|---|
| 基础模型 | Qwen/Qwen3.6-27B |
| 量化 | MXFP4(4-bit,group_size=32,packed uint32 + E8M0 scales) |
| 视觉塔 / projector | fp16(不量化) |
| MTP 层数 | 1(quantization_config.mtp_num_hidden_layers: 1) |
| 参数量 | ~27B |
| 上下文长度 | 262,144 |
| 架构 | Qwen3.5(Gated DeltaNet 混合),VL 变体 |
| 文件大小 | ~16GB(4 个 safetensors 分片) |
| 工具链 | junainfinity/mlx-mtp,--mode mxfp4 --verify |
| 部分 | 精度 | 说明 |
|---|---|---|
| 语言模型主干 | MXFP4(4-bit,group_size=32) | packed uint32 权重 + uint8 E8M0 scales |
| 视觉塔 / projector / merger | fp16 | 完整保留图像能力 |
MTP 头(mtp.*) | 保留(fp16) | 自起草投机解码 |
| SSM 敏感参数 | 高精度 | a_log、dt_bias、conv1d、ssm_* |
Qwen3.6-27B-MXFP4-VL-MTP/
├── config.json # mlx_mtp 元信息 + quantization_config.mtp_num_hidden_layers: 1
├── generation_config.json
├── preprocessor_config.json # 图像预处理配置
├── chat_template.jinja
├── tokenizer.json
├── tokenizer_config.json
├── vocab.json
├── merges.txt
├── model.safetensors.index.json
├── model-00001-of-00004.safetensors # ~5.3GB
├── model-00002-of-00004.safetensors # ~5.3GB
├── model-00003-of-00004.safetensors # ~5.4GB
└── model-00004-of-00004.safetensors # ~0.2GB
注意:MTP 是否生效由
config.json中的quantization_config.mtp_num_hidden_layers控制(1= 开启,0或缺失 = 关闭)。
mlx-mtp(Python,推荐):使用 junainfinity/mlx-mtp 自带的推理引擎和 OpenAI 兼容服务(mlx-mtp-serve)。
mlx-swift-lm(Swift):通过支持 Qwen3.5/3.6 VL 的 mlx-swift-lm 的 VLMModelFactory 直接加载本仓库/目录。
mlx-mtp-quantize \
--src Qwen/Qwen3.6-27B-bf16 \
--out Qwen3.6-27B-MXFP4-VL-MTP \
--mode mxfp4 \
--verify
BF16 源模型约 55GB → MXFP4 约 16GB。视觉塔、MTP 头和 SSM 敏感参数由工具的 skip 规则排除在量化之外。
mtp_num_hidden_layers > 0(如 mlx-mtp,或带 Qwen3.5 MTP 支持的 mlx-swift-lm 构建)Apache-2.0,与基座模型一致。按原样提供,不作任何担保。