PULSE-2B

PULSE-2B is a fully standalone, merged, and 4-bit quantized version of Google's Gemma-2 2B Instruct.

It combines the base weights with a custom LoRA fine-tuning adapter into a single, high-performance weight set optimized for fast on-device inference using MLX on Apple Silicon (M1/M2/M3/M4).

Key Highlights

• Standalone: No need to download base weights or attach adapters separately.
• Quantization: 4-bit group-quantized (group_size=64) via MLX (~1.54 GB VRAM).
• Inference Speed: ~57 tokens/sec on Apple Silicon.
• Training Base: Fine-tuned over 600 iterations from google/gemma-2-2b-it.

---

🚀 Quickstart Guide

1. Installation

Ensure you are running macOS on Apple Silicon:

pip install -U mlx-lm

2. Command Line Interface (CLI)

Generate text instantly in your terminal:

mlx_lm.generate \
  --model muonai/PULSE-2B \
  --prompt "Instruction:\nExplain the concept of fine-tuning in 2 simple sentences.\n\nResponse:\n" \
  --max-tokens 150

Start an interactive chat session:

mlx_lm.chat --model muonai/PULSE-2B

3. Python API (mlx_lm)

from mlx_lm import load, generate

model_id = "muonai/PULSE-2B"

# Load standalone model and tokenizer
model, tokenizer = load(model_id)

prompt = "Instruction:\nGive 3 quick tips for daily productivity.\n\nResponse:\n"

response = generate(
    model, 
    tokenizer, 
    prompt=prompt, 
    max_tokens=200, 
    verbose=True
)

print(response)

---

📊 Performance Benchmark

Tested on Apple Silicon locally using mlx_lm.generate:

Metric	Measurement
Generation Speed	~56.97 tokens/sec
Prompt Processing	~49.83 tokens/sec
Peak Memory Usage	~1.54 GB
Precision	~4.50 bits/weight

---

📜 License

This model is a derivative work based on Google's Gemma 2 weights and is subject to the official Gemma Terms of Use.