Choosing a model by parameter count or "what fits" often misleads: the biggest model that fits VRAM can be slower or lower-quality than a newer, better-quantized alternative. whichllm focuses on the pragmatic question many engineers actually face — which local LLM will run and give the best real-world performance on my hardware?
What Sets It Apart
- Evidence-first ranking: merges multiple benchmark sources (LiveBench, Artificial Analysis, Aider, Chatbot Arena ELO, Open LLM Leaderboard) and tags evidence by confidence so scores reflect data quality, not just size. Recency-aware lineage demotion prevents stale leaderboards from overranking older models.
- Hardware-aware fit and speed: models are evaluated by estimated VRAM (weights + KV cache + activations + overhead), backend/quantization efficiency, and bandwidth-bound token/sec projections. Simulation mode (--gpu) lets you mimic other cards for planning.
- Practical output and automation: one-command flow to pick, download, and run a model; JSON output for scripting; snippets for local runtimes (llama-cpp-python, AWQ/GPTQ, GGUF). Live HuggingFace queries with cached fallbacks balance freshness and reliability.
Who It's For (and Trade-offs)
Great fit if you need to pick or plan hardware (buying a GPU), want a quick, realistic local-model recommendation for a given machine, or need scriptable JSON output to wire into deployment workflows. It’s also useful for developers who want a reproducible, evidence-graded model shortlist rather than manual trial-and-error.
Look elsewhere if you require exhaustive, lab-grade microbenchmarks for a custom build (whichllm provides planning-range estimates, not per-variant lab runs), need a GUI-first experience, or rely entirely on cloud-hosted managed models. Note: rankings depend on available benchmark sources and HuggingFace metadata; speed figures are planning estimates with confidence ranges rather than absolute guarantees.
