whichllm is a command-line tool that recommends the best-performing local language model for your hardware by auto-detecting your GPU, CPU, and RAM, then ranking candidates from HuggingFace by real benchmark scores rather than parameter count alone.
The problem it solves is that fitting a model into available VRAM is straightforward, but choosing which runnable model actually performs best is not. whichllm addresses this by ranking models on recency-aware benchmarks and real inference speed measurements on your specific hardware, rather than relying on model size as a proxy for quality. It estimates generation speed in tokens per second and quality scores, then surfaces the model that balances both metrics best for your system. The tool supports multiple fitting strategies: by default it considers models that fit entirely in GPU VRAM, use partial RAM offload, or run on CPU when they appear usable; users can restrict to full-GPU fits only or adjust headroom thresholds for more conservative recommendations.
The tool suits developers and researchers who want to run local models without guesswork about which candidate will actually perform well on their hardware. It is designed for one-off use without installation, though it can be installed for frequent use. The README emphasizes that whichllm differs from size-only selection tools by ranking a 27B model higher than a 32B model when benchmarks show better quality and recency, even though both fit the same card. Users can simulate hardware before purchasing by specifying a GPU model, and the tool tracks live HuggingFace data so recommendations reflect current model availability rather than static snapshots.
Development activity shows consistent engagement with the project's core purpose. The tool maintains live tracking of HuggingFace model data to keep recommendations current. Speed estimates are marked with confidence indicators to signal when they are approximate versus measured. The project includes a Japanese-language README alongside the English version, indicating attention to international users.
