Ollama and llama.cpp are great starting points for local LLMs, but once models become part of a real workflow, more specialized runtimes become necessary. vLLM offers high-throughput inference with PagedAttention, continuous batching, and OpenAI-compatible APIs for serving models as backend infrastructure. SGLang targets structured generation and agent-style workloads with prefix caching and constrained outputs. For Apple Silicon, vMLX provides a Mac-native serving stack built on Apple's MLX framework. MLC-LLM enables deployment to browsers (WebGPU), iOS, and Android, while ExLlamaV3 squeezes performance from consumer GPUs with its EXL3 quantization format. Other notable tools include llama-swap for routing, TensorRT-LLM for Nvidia, LMDeploy, Lemonade for AMD, KTransformers for heterogeneous CPU/GPU inference, and LocalAI.

8m read timeFrom xda-developers.com
Post cover image
Table of contents
vLLM and SGLang turn local models into infrastructurevMLX gives Macs a more serious local appMLC-LLM and ExLlamaV3 target specific hardware problemsThere's more than just Ollama and llama.cpp out there
57 Impressions