richardyoung/Kimi-K2-Instruct-0905-MLX-3bit

🌙 Kimi K2 Instruct - MLX 3-bit

State-of-the-Art 671B MoE Model, Optimized for Apple Silicon

Original Model | MLX Framework | More Quantizations

---

📖 What is This?

This is a compact 3-bit quantized version of Kimi K2 Instruct, optimized to run on Apple Silicon (M1/M2/M3/M4) Macs using the MLX framework. The 3-bit version offers aggressive compression for fast inference with reasonable quality - great for testing and resource-constrained scenarios!

✨ Why You'll Love It

• 🚀 Massive Context Window - Handle up to 262,144 tokens (~200,000 words!)
• 🧠 671B Parameters - One of the most capable open models available
• ⚡ Apple Silicon Native - Fully optimized for M-series chips with Metal acceleration
• 💨 Compact Size - Only ~420 GB, very space-efficient
• 🏃 Fast Inference - Quick generation speeds
• 🌏 Bilingual - Fluent in both English and Chinese
• 💬 Instruction-Tuned - Ready for conversations, coding, analysis, and more

🎯 Quick Start

Installation

pip install mlx-lm

Your First Generation (3 lines of code!)

from mlx_lm import load, generate

model, tokenizer = load("richardyoung/Kimi-K2-Instruct-0905-MLX-3bit")
print(generate(model, tokenizer, prompt="Explain quantum entanglement simply:", max_tokens=200))

That's it! 🎉

💻 System Requirements

Component	Minimum	Recommended
Mac	M1 or newer	M2 Ultra / M3 Max / M4 Max+
Memory	40 GB unified	80 GB+ unified
Storage	500 GB free	Fast SSD (1+ TB)
macOS	12.0+	Latest version

⚡ Note: The 3-bit version is very memory-efficient and runs fast on most Apple Silicon Macs!

📚 Usage Examples

Command Line Interface

mlx_lm.generate \
  --model richardyoung/Kimi-K2-Instruct-0905-MLX-3bit \
  --prompt "Write a Python script to analyze CSV files." \
  --max-tokens 500

Chat Conversation

from mlx_lm import load, generate

model, tokenizer = load("richardyoung/Kimi-K2-Instruct-0905-MLX-3bit")

conversation = """<|im_start|>system
You are a helpful AI assistant specialized in coding and problem-solving.<|im_end|>
<|im_start|>user
Can you help me optimize this Python code?<|im_end|>
<|im_start|>assistant
"""

response = generate(model, tokenizer, prompt=conversation, max_tokens=500)
print(response)

Advanced: Streaming Output

from mlx_lm import load, generate

model, tokenizer = load("richardyoung/Kimi-K2-Instruct-0905-MLX-3bit")

for token in generate(
    model,
    tokenizer,
    prompt="Tell me about the future of AI:",
    max_tokens=500,
    stream=True
):
    print(token, end="", flush=True)

🏗️ Architecture Highlights

Click to expand technical details

Model Specifications

Feature	Value
Total Parameters	~671 Billion
Architecture	DeepSeek V3 (MoE)
Experts	384 routed + 1 shared
Active Experts	8 per token
Hidden Size	7168
Layers	61
Heads	56
Context Length	262,144 tokens
Quantization	3.502 bits per weight

Advanced Features

• 🎯 YaRN Rope Scaling - 64x factor for extended context
• 🗜️ KV Compression - LoRA-based (rank 512)
• ⚡ Query Compression - Q-LoRA (rank 1536)
• 🧮 MoE Routing - Top-8 expert selection with sigmoid scoring
• 🔧 FP8 Training - Pre-quantized with e4m3 precision

🎨 Other Quantization Options

Choose the right balance for your needs:

Quantization	Size	Quality	Speed	Best For
8-bit	~1 TB	⭐⭐⭐⭐⭐	⭐⭐⭐	Production, best quality
6-bit	~800 GB	⭐⭐⭐⭐	⭐⭐⭐⭐	Sweet spot for most users
5-bit	~660 GB	⭐⭐⭐⭐	⭐⭐⭐⭐	Great quality/size balance
4-bit	~540 GB	⭐⭐⭐	⭐⭐⭐⭐⭐	Faster inference
3-bit (you are here)	~420 GB	⭐⭐	⭐⭐⭐⭐⭐	Very fast, compact
2-bit	~320 GB	⭐⭐	⭐⭐⭐⭐⭐	Fastest, most compact
Original	~5 TB	⭐⭐⭐⭐⭐	⭐⭐	Research only

🔧 How It Was Made

This model was quantized using MLX's built-in quantization:

mlx_lm.convert \
  --hf-path moonshotai/Kimi-K2-Instruct-0905 \
  --mlx-path Kimi-K2-Instruct-0905-MLX-3bit \
  -q --q-bits 3 \
  --trust-remote-code

Result: 3.502 bits per weight (includes metadata overhead)

⚡ Performance Tips

Getting the best performance

1. Close other applications - Free up as much RAM as possible
2. Use an external SSD - If your internal drive is full
3. Monitor memory - Watch Activity Monitor during inference
4. Adjust batch size - If you get OOM errors, reduce max_tokens
5. Keep your Mac cool - Good airflow helps maintain peak performance
6. Great for speed - Perfect for rapid iteration and testing

⚠️ Known Limitations

• 🍎 Apple Silicon Only - Won't work on Intel Macs or NVIDIA GPUs
• 💾 Quality Trade-off - 3-bit quantization impacts quality noticeably
• 🎯 Best for Testing - Consider 4-bit or higher for production
• 📊 Experimental Use - May produce less accurate outputs than higher bits
• 🌐 Bilingual Focus - Optimized for English and Chinese
• 💡 Use Case - Great for prototyping, testing, and resource-constrained environments

💡 Recommendation: For production use, consider the 5-bit, 6-bit, or 8-bit versions for better quality!

📄 License

Apache 2.0 - Same as the original model. Free for commercial use!

🙏 Acknowledgments

• Original Model: Moonshot AI for creating Kimi K2
• Framework: Apple's MLX team for the amazing framework
• Inspiration: DeepSeek V3 architecture

📚 Citation

If you use this model in your research or product, please cite:

@misc{kimi-k2-2025,
  title={Kimi K2: Advancing Long-Context Language Models},
  author={Moonshot AI},
  year={2025},
  url={https://huggingface.co/moonshotai/Kimi-K2-Instruct-0905}
}

🔗 Useful Links

• 📦 Original Model: moonshotai/Kimi-K2-Instruct-0905
• 🛠️ MLX Framework: GitHub
• 📖 MLX LM Docs: GitHub
• 💬 Discussions: Ask questions here!

---

Quantized with ❤️ by richardyoung

If you find this useful, please ⭐ star the repo and share with others!

Created: October 2025 | Format: MLX 3-bit