Irfanuruchi/Phi-2-MLX-5bit

Phi-2 MLX 5-bit

This repository provides a 5-bit MLX-quantized version of Microsoft Phi-2, optimized for higher output quality while remaining suitable for local, offline inference on Apple Silicon.

This variant offers better instruction-following and coherence compared to the 4-bit version, at a modest increase in memory usage.

---

Model Details

• Base model: microsoft/phi-2
• Architecture: Decoder-only Transformer
• License: MIT
• Quantization: MLX static quantization (≈5.5 bits per weight)
• Target hardware: Apple Silicon (M1 / M2 / M3)

---

Performance Characteristics

Metric	Value
Disk size	~1.9–2.1 GB
Peak RAM usage	~2.0–2.2 GB
Inference speed	Moderate
Instruction quality	Higher

---

Usage

mlx_lm.generate \
  --model /path/to/Phi-2-MLX-5bit \
  --prompt "Explain the FFT in simple terms." \
  --max-tokens 120

---

Notes

• This is a quantized conversion, not a fine-tuned model.
• The 5-bit version is recommended for:
  • better reasoning consistency
  • fewer repetitions
  • improved instruction adherence
• For maximum speed and lower memory usage, see the 4-bit variant.

---

License

This repository redistributes a quantized MLX conversion of Microsoft Phi-2.

• Original model license: MIT
• MLX conversion: MIT

See LICENSE for details.