PowerInfer

Reading on inference

• Choosing an inference runtimeThe wrong runtime costs you throughput you never knew you were leaving on the table. The right one is usually decided by traffic shape, not model size.
• Context is a budget, not a containerThe instinct is to keep everything in case it is needed. The models are not filing systems, and past a certain length the middle of your context is doing nothing but costing money.
• Choosing a model is a routing problemThere is no best model. There is a best model for a task, at a quality bar, within a latency and cost budget — and that decision is a routing policy, not a benchmark.

• llama.cppPortable CPU/GPU inference for quantized GGUF models.
• vLLMPaged-attention inference engine with an OpenAI-compatible server.

Also in inference

• vLLMvLLMPaged-attention inference engine with an OpenAI-compatible server.Use when General GPU serving under bursty, high-concurrency traffic.Skip when Traffic is steady and low-concurrency, so batches never fill.Apache-2.0 · self-hosted · free
• SGLangSGLangStructured generation runtime built on RadixAttention prefix reuse.Use when A large share of requests share a long prefix, or you need constrained decoding.Skip when You want the fewest moving parts and vLLM already clears your bar.Apache-2.0 · self-hosted · free
• llama.cppllama.cppPortable CPU/GPU inference for quantized GGUF models.Use when The constraint is hardware, not throughput — no GPU, or an edge box.Skip when You need maximum concurrent throughput on datacentre GPUs.MIT · self-hosted · free
• OllamaOllamaLocal model runner with a single-binary distribution and HTTP API.Use when Local model access for a developer or small team, with no setup.Skip when You need production throughput or a stable server API.MIT · self-hosted · free
• TensorRT-LLMTensorRT-LLMNVIDIA-optimized inference with inflight batching and FP8.Use when You are committed to NVIDIA hardware and need the last of the throughput.Skip when Hardware portability matters, or you have no GPUs to tune against.Apache-2.0 · self-hosted · free
• Text Generation InferenceText Generation InferenceHugging Face production server for LLMs with tensor parallelism.Use when Multi-GPU tensor parallelism inside an existing Hugging Face stack.Skip when The model fits on one GPU, where this is overkill.Apache-2.0 · self-hosted · free
• LM StudioLM StudioDesktop app for running and serving local models with an OpenAI API.Use when A desktop GUI for running and serving local models, aimed at one operator.Skip when You need headless, multi-tenant deployment.proprietary · self-hosted · free
• Triton Inference ServerTriton Inference ServerMulti-framework inference server for ONNX, TensorRT and Python.Use when One endpoint serving ONNX, TensorRT and Python models side by side.Skip when You serve one model family and want peak LLM throughput.BSD-3-Clause · self-hosted · free
• llamafilellamafilePackages a model and its runtime into a single executable.Use when Shipping a model as a single executable with no install step.Skip when You need batching, multi-GPU serving or a production HTTP server.Apache-2.0 · self-hosted · free
• KoboldCppKoboldCppGGUF inference with a focused text-adventure UI and API.Use when Local generation with a UI tuned for interactive, long-form use.Skip when You need headless serving at scale.MIT · self-hosted · free
• MarlinMarlinQuantised GEMM kernels for near-GPU speed at 4-bit.Use when Near-GPU throughput from 4-bit weights, as a kernel inside another runtime.Skip when You need a scheduler and a server as well as a kernel.Apache-2.0 · self-hosted · free