ds4: antirez's New Inference Engine — 7.1k Stars in 4 Days

Описание: ds4 (DwarfStar 4) is a specialized local inference engine for DeepSeek V4 Flash, written in C by Salvatore Sanfilippo — antirez, the creator of Redis. The repo is 4 days old and already past 7,000 GitHub stars. Metal on macOS, CUDA on Linux. Built around the architectural bet that the KV cache should be a first-class disk citizen, not a RAM resident. Runs DeepSeek V4 Flash on 128GB MacBooks via a special asymmetric 2-bit quant (only routed MoE experts quantized). Tool/function calling with OpenAI + Anthropic compatibility, thinking mode, MTP speculative decoding, 1M-token context window with disk-backed KV cache. Intentionally narrow — not a generic GGUF runner, not a wrapper, not a framework. MIT.

0:00 7.1k stars in 4 days
0:16 Walking the GitHub repo
1:17 What ds4 actually is — and isn't
2:21 Why DeepSeek V4 Flash specifically
3:23 Install + first run
4:31 The download + server commands
4:58 KV cache as a first-class disk citizen
6:08 By the numbers
7:11 Antirez on AI-assisted development
8:09 Wrap

Repo: https://github.com/antirez/ds4
Weights: https://huggingface.co/antirez/deepse...

— Prism Labs · @prismlabsai