DeepSeek researchers have open-sourced Nano-VLLM, a minimalist implementation of the popular VLLM (LLM serving system) built from scratch with three key optimizations:

1. 5x Smaller Memory Footprint (under 2GB for 7B models)

• Achieved via dynamic page-aware KV cache compression
• Supports CPU offloading for edge devices

2. Python-First Architecture

• No CUDA dependencies (unlike original VLLM)
• Pure NumPy backend for prototyping

3. Experimental Features

• Speculative token prefill - 30% faster first token
• BitDelta quantization - 3-bit weights with <1% perplexity loss

Currently supports LLaMA3 and DeepSeek's own 7B models. Benchmark shows 23 tokens/sec on consumer CPUs (i7-13700K).