This model was converted to MLX format from mlx-community/Qwen3.8-27B-bf16 using mlx-vlm version 0.6.13. Refer to the original model card for more details on the model.

Quick use with mlx

pip install -U mlx-vlm

python -m mlx_vlm.generate --model zherebetskyy/Qwen3.8-27B-6bit-mlx --max-tokens 100 --temperature 0.0 --prompt "Describe this image." --image <path_to_image>

Details

What's this?

Qwen3.8-27B is an advanced 27B-parameter dense model from Alibaba Cloud. It features a hybrid linear/full attention architecture (3:1 ratio across 64 layers) that merges highly efficient DeltaNet-style linear attention with standard full softmax attention layers. It natively supports a 262K context window, multimodal vision, video parsing, and multi-token prediction (MTP).

This repository hosts a 6-bit linear affine group-quantized conversion (group size 64) optimized specifically for Apple Silicon hardware. Averaging roughly 6.6 bits/weight, this model card layout compresses the 27B structure to fit comfortably into 21–23 GB of memory—leaving ideal unified memory headroom for long-context execution and media workloads on 64GB Mac systems.

🎛️ Architecture Details

Spec	Value
Total Parameters	27.8B (dense, all active)
Layer Configuration	64 layers (3x linear attention + 1x full attention, 16 repetitions)
Standard Attention	24 Q heads, 4 KV heads (GQA), head_dim 256
Linear Attention	16 QK heads, 48 V heads, head_dim 128
Context Length	262K native, up to 1M+ using YaRN scales
Vocabulary Size	248K tokens
Native Modalities	Text, Image, Video
Model Type Tag	`qwen3_8`

Conversion Details

Source Weights: mlx-community/Qwen3.8-27B-bf16 (extracted from the raw FP16/BF16 base distribution)
Quantization Format: Uniform MLX Group Affine 6-bit
Vision Tower Precision: 100% retained. The vision tower parameters remain unquantized to guarantee maximum feature parsing and image description capability.
Tool Calling Support: Enhanced via a fixed Jinja template that patches runtime exceptions on C++ execution engines (LM Studio, MLX Local Server) and correctly handles developer system roles.

⚡ Quick Start

For fast, isolated package handling without polluting your system Python env, we recommend using uv (a blazing-fast installer written in Rust).

1. Unified Command Line (CLI)

Multimodal Image Prompt (Vision)

bash

uv run --with mlx-vlm mlx_vlm.generate \
  --model zherebetskyy/Qwen3.8-27B-6bit-mlx \
  --image /path/to/your_image.jpg \
  --prompt "Describe the layout and contents of this image in detail."

Standard Text Generation

bash

uv run --with mlx-lm mlx_lm.generate \
  --model zherebetskyy/Qwen3.8-27B-6bit-mlx \
  --prompt "Write a high-performance Python script using asyncio."

2. Python API Core Examples

📸 Multimodal Vision Interaction

python

from mlx_vlm import load, generate
from mlx_vlm.prompt_utils import apply_chat_template

# Load the model and its matching multi-modal visual processor
model, processor = load("zherebetskyy/Qwen3.8-27B-6bit-mlx")

image = ["/path/to/your_image.jpg"]
prompt = "Analyze the spatial layout of this scene."

# Wrap prompt and images inside the chat format
formatted = apply_chat_template(processor, model.config, prompt, num_images=len(image))

# Generate visual outputs on the Apple Silicon GPU
result = generate(model, processor, formatted, image, max_tokens=400, temp=0.7)
print(result.text)

✍️ Standard Text Generation

python

from mlx_lm import load, generate

model, tokenizer = load("zherebetskyy/Qwen3.8-27B-6bit-mlx")

response = generate(
    model, 
    tokenizer, 
    prompt="Design a clean multi-tenant architecture for a SaaS app.", 
    max_tokens=500, 
    temp=0.6
)
print(response)

⚠️ Crucial System Prompt Note

To prevent the model from underperforming, the first line of your user/system system prompt must conform to the following explicit format:

You are Qwen, created by Alibaba Cloud. You are a helpful assistant.

You can seamlessly append your customized contextual instructions or agent boundaries right after this line.

🧠 Thinking Toggle Mode

The included custom Jinja template natively reads and extracts <|think_on|> and <|think_off|> formatting tokens directly from message blocks to flag the model's internal step-by-step reasoning system on the fly:

text

System: You are an engineering assistant. <|think_off|>
User: What is the optimal time complexity of a binary search tree?

(Triggers a direct response without outputting verbose reasoning paths).

text

System: You are a cryptography expert. <|think_on|>
User: Implement a side-channel attack resilient RSA algorithm block in C.

(Forces the model to brainstorm structural safety vectors step-by-step before streaming code).

📊 Recommended Sampling Parameters

Ensure you allocate over 128K+ of your context window if utilizing active reasoning (thinking) configurations.

Mode	Temperature (`temp`)	Top-P	Top-K	Min-P	Repeat Penalty	Presence Penalty
💻 Thinking (Coding/Logic)	`0.6`	`0.95`	`20`	`0`	`1.0` (Off)	Off
🌐 Thinking (General Tasks)	`1.0`	`0.95`	`20`	`0`	`1.0` (Off)	`1.5`
⚡ Fast Generation (No-Thinking)	`0.7`	`0.80`	`20`	`0`	`1.0` (Off)	`1.5`

🔗 Reference Links

Original Architecture Baseline: Qwen/Qwen3.8-27B
Quantization Origin Source: mlx-community/Qwen3.8-27B-bf16
Official Model Announcement: Alibaba Qwen Team Blog

👤 Credits & Authorship

Model Creators: Alibaba Cloud (Qwen Team)
MLX 6-bit Conversion Specialist: zherebetskyy