Favicon of whichllm

whichllm

A local LLM recommendation tool that ranks HuggingFace models for your hardware, estimates speed, and compares GPU upgrades.

whichllm is a command-line tool for people choosing models to run on their own hardware. It checks your machine and recommends HuggingFace models based on benchmark results, memory fit, and estimated generation speed.

  • Benchmark rankings. Compare models using evidence from LiveBench, Artificial Analysis, Aider, and other evaluations, with discounts for older results and less reliable claims.
  • Hardware fit. See which models can run entirely on your GPU, split across GPU and system memory, or use only the CPU, with filters for speed and available memory.
  • Task selection. Narrow recommendations to coding, vision, math, or general use, and identify complete GGUF models already installed in LM Studio.
  • Upgrade planning. Simulate single or multiple GPUs, compare upgrade candidates, or find hardware suitable for a specific model before buying.
  • Local chat and code. Download a model and start a local chat, generate Python snippets, or export recommendations as JSON and Markdown.

The tool is open source under the MIT license. It detects NVIDIA, AMD, Intel, and Apple Silicon hardware, as well as CPU-only systems. Local inference supports GGUF through llama-cpp-python and FP16/BF16 through Transformers; AWQ/GPTQ support is available on Linux with NVIDIA hardware. Apple Silicon and CPU-only recommendations are restricted to GGUF.

Live model listings and benchmark updates use external services, including HuggingFace. Offline recommendations use cached data or curated fallback catalogs. Speed figures are estimates rather than measurements on your machine, and default recommendations can include models near memory limits or using partial CPU offload.

Similar to whichllm