whichllm 命令

常用示例

说明

whichllm detects local hardware (GPU model and VRAM, CPU, RAM, OS) and ranks open-weight large language models from HuggingFace and Ollama by how well they will actually run on that machine. Instead of treating "fits in VRAM" as the only criterion, it combines a fit check with recency-aware benchmark scores from sources such as LiveBench, Artificial Analysis, Aider, and the Chatbot Arena ELO leaderboard, and applies penalties for quantization, partial offload, and MoE architectures. The tool is designed for the common practical question "which model should I download tonight" rather than for marketing claims. The default invocation prints a short ranked table; subcommands extend the same engine to launch interactive sessions, plan hardware upgrades, or emit code snippets for direct integration.

参数

hardware

Print detected GPU, CPU, RAM, and OS information without ranking models.

run _model_

Download _model_ via Ollama and start an interactive chat session.

plan _model_

Reverse lookup: estimate which GPU or RAM tier is needed to run _model_ at usable speed.

snippet _model_

Print a ready-to-paste Python snippet that loads _model_ from HuggingFace or Ollama.

--gpu _name_

Override hardware detection and rank as if running on the named GPU (e.g. "RTX 4090").

--cpu-only

Restrict ranking to models that run acceptably without a GPU.

--top _N_

Show the top _N_ ranked models instead of the default short list.

--quant _type_

Filter results by quantization (e.g. _Q4_K_M_, _Q5_K_M_, _Q8_0_, _fp16_).

--profile _use_case_

Bias ranking towards a specific profile (_coding_, _vision_, _math_, _general_).

--json

Emit machine-readable JSON instead of the formatted table.

--refresh

Bypass the local cache and refetch benchmark data.

--version

Print version and exit.

--help

Print help and exit.