A 4-bit MLX conversion of a Qwen 3.6 35B MoE (Mixture-of-Experts) vision-language model, fine-tuned with Unsloth for code generation and step-by-step reasoning, distilled from DeepSeek v4 Pro reasoning traces.

Optimized for Apple Silicon (M1/M2/M3/M4) via the MLX framework.

Model Highlights

Architecture: Qwen3.5 MoE (Qwen3_5MoeForConditionalGeneration) — sparse Mixture-of-Experts, only a subset of parameters is active per token → fast inference for its total size
Total parameters: ~35B (fewer active per forward pass thanks to MoE routing)
Quantization: 4-bit affine (group size 64), with sensitive MoE router/gate layers kept at 8-bit to preserve routing quality (Unsloth-style mixed precision)
Multimodal: supports image and video inputs (<|vision_start|> / <|image_pad|> / <|video_pad|> tokens)
Tool calling: native function/tool-calling support in the chat template
Training: distilled from DeepSeek v4 Pro reasoning outputs, focused on coding and chain-of-thought reasoning tasks
Language: English

Use with mlx

pip install -U mlx-lm

python

from mlx_lm import load, generate

model, tokenizer = load("10Hen10/Qwen3.6-35B-unsloth-code-reasoning-deepseek-v4-pro-destilled-Mlx-4bit")

prompt = "Write a Python function that checks whether a binary tree is balanced, and explain your reasoning."

messages = [{"role": "user", "content": prompt}]
prompt = tokenizer.apply_chat_template(messages, add_generation_prompt=True)

response = generate(model, tokenizer, prompt=prompt, verbose=True, max_tokens=1024)

Or from the command line:

bash

mlx_lm.generate --model 10Hen10/Qwen3.6-35B-unsloth-code-reasoning-deepseek-v4-pro-destilled-Mlx-4bit \
  --prompt "Explain the difference between a mutex and a semaphore." \
  --max-tokens 512

Memory requirements

Approximate unified-memory usage on Apple Silicon:

Component	Approx. size
Model weights (4-bit)	~20–22 GB
KV cache + overhead	~2–4 GB
Recommended RAM	32 GB+ (comfortable on 36/48/64 GB machines)

Intended use

Code generation, completion and explanation
Reasoning-heavy tasks (math, logic, debugging)
Agentic workflows with tool calling
Local, private inference on macOS

Limitations

This is a community distillation — no official benchmark results are provided; evaluate on your own tasks before production use.
4-bit quantization may degrade quality on nuanced tasks compared to full-precision weights.
Reasoning style and knowledge are inherited from the teacher model and the distillation dataset; the model may hallucinate or produce incorrect code.
English-focused; other languages may work but are not the training target.

Acknowledgements

Qwen for the base model architecture
Unsloth for efficient fine-tuning
DeepSeek for the teacher reasoning traces
Apple MLX for the inference framework