I got GPT-2 running on a $3 Arduino. No cloud. No subscription. Just quantization.

The Problem:
Local LLMs are great until you try to run them on real hardware. GPT-2 takes 500MB+ just for the KV cache. On an embedded device? Forget it.

The Solution: KVQuant
I compressed the KV cache from full precision to 1-bit per value using per-channel symmetric quantization. Mixed INT8 for attention scores where precision matters more.

Results:

• 3.2x faster inference
• 73% memory reduction
• Runs on ESP32-class hardware

Code:

from kvquant import QuantizedModel
model = QuantizedModel("gpt2", bits=1)
model.generate("Hello world")

Benchmark:
| Model | Memory | Latency |
|-------|--------|---------|
| FP16 GPT-2 | 520MB | 2.1s |
| KVQuant-1b | 140MB | 0.65s |

GitHub: https://github.com/AmSach/kvquant

This isn't a demo — it's a real quantization library with INT8 kernels and hardware-aware optimizations. Pull the repo, run the examples, see for yourself.