Rondine is an open-source control plane for local LLMs. It detects RAM and VRAM, recommends models that fit, applies hardware-tuned settings, downloads weights, and starts an OpenAI-compatible server. It supports Apple Silicon, NVIDIA GPUs, and DGX Spark through llama.cpp, MLX-LM, and vLLM. Instead of creating another inference engine, Rondine coordinates proven runtimes and shows every launch plan before execution.

Hi Product Hunt! I built Rondine after repeatedly spending more time configuring local inference than actually using it. Running a local LLM involves many interconnected choices: model, quantization, context size, inference engine, GPU offloading, KV cache, and memory limits. A configuration that works well on an Apple Silicon Mac may be completely wrong for an NVIDIA workstation or DGX Spark. Rondine turns those decisions into a hardware-aware workflow. It inspects the machine, recommends models that fit, produces a reviewable configuration, downloads the weights, and launches an OpenAI-compatible endpoint using llama.cpp, MLX-LM, or vLLM. Rondine means “swallow” in Italian, a tiny bird helping suspiciously large models take flight. I’d love to hear which hardware and local models you use, and where setup still causes the most friction.

Running local models is becoming much more interesting. Curious where you've seen the strongest pull so far—privacy, latency, cost, or something you didn't expect?

Rondine

Run the right local LLM for your hardware

Run the right local LLM for your hardware