llm A field note by Vikrant Sharma
The Redis creator just shipped a local LLM runner
Antirez released ds4, a tool for running LLMs locally without the usual Python dependency chaos. It is written in C.
Salvatore Sanfilippo, who built Redis, just released ds4, a tool for running LLMs on your machine without Docker or virtual environments. It is written in C. The pitch is simple. Most local LLM setups involve Python, conda environments, version mismatches, and eventually a Stack Overflow search at 2am. ds4 compiles to a single binary. You download a model, point ds4 at it, done. What caught my attention is the creator. Antirez does not ship hobby projects. Redis powered half the internet’s caching layer for a decade. If he thinks local LLM tooling is broken enough to warrant a rewrite in C, he is probably right. The timing is interesting. OpenAI’s API costs dropped, but latency and rate limits still push people toward local inference for prototyping. Tools like Ollama exist, but they abstract too much for anyone who wants control over quantisation or context windows. ds4 sits in the middle: less hand-holding than Ollama, less dependency hell than llama.cpp’s Python bindings. I have not tested it yet. The repo went live two days ago. But the approach matches what I needed six months ago when debugging a RAG pipeline. I wanted to swap models fast, check perplexity on different quantisations, and not reinstall CUDA drivers every other day. ds4’s README suggests it handles that. One risk: writing LLM inference in C means you own the memory bugs. Rust would have been safer. But C compiles everywhere, and Antirez knows what he is doing with manual allocation. Redis ran in production for years without segfaulting. I will try it this week on a Phi-3 model and see if the ergonomics hold up. If it does what the docs claim, it replaces three tools in my workflow.
