Core conclusion

Another native AI utility breaks the mold. WhichLLM is a command line tool: enter your graphics card model, it will automatically pull data from HuggingFace, combine it with real benchmark scores such as LiveBench, Aider, Chatbot Arena, etc., and rank you according to the dual dimensions of "Hardware Adaptation × Model Quality".

This news is useful for WayToClawEarn users who often run local models: choosing a large model is no longer a guesswork, but a decision supported by data.

Key Points

• Release time: 2026-05-15
• Release form: GitHub open source (MIT license), PyPI / Homebrew installation
• Installation method: pipx install whichllm or brew install whichllm
• Core functions: Automatically detect hardware → Filter models by VRAM → Rank by benchmark scores → Run with one click
• Current Star Count: GitHub 5000+

Background: The real pain points of local model selection

Since the intensive release of open source models such as DeepSeek V4, Qwen3, and Llama 4, more and more developers and content creators have chosen to run LLM locally. But a core question always remains: With so many models, which one is best for my machine? **

The traditional approach is to use Ollama to manually download and test one by one, or to make a rough estimate based on the number of model parameters. However, both methods have obvious flaws - a large number of model parameters does not mean good inference quality, and the time cost of trying each one is very high.

WhichLLM founder Andyyyy64 shared this project on HN: It is not another simple tool of "what can run", but a scoring engine of "what is worth running".

SEO: Local large model selection, WhichLLM, GGUF model ranking, RTX 4090 best model GEO: starting with TL;DR, precise numbers are extra points

Core function disassembly

Genuine recommendations from WhichLLM

Recommended by graphics card

The special value of the MoE model

It is worth special mentioning that WhichLLM has specially optimized the MoE (Mixed Experts) model. For example, Qwen3-30B-A3B has a total score of 82.7 but an inference speed of up to 102 t/s - because only 3B parameters are activated for each inference, it is close to the 30B model in quality. This allows RTX 4060 users with 8GB VRAM to run well.

Practical value for AI content creators

Optimal use cases for WhichLLM include:

1. Do your homework before buying a graphics card: Use whichllm --gpu "RTX 5090" to simulate future configurations to avoid blind purchases
2. Reduce API costs: Running the Qwen3.6-27B level model locally, the average monthly cost is about 10-20% of that of the cloud API
3. Offline workflow: Cooperating with local Agent frameworks (such as OpenClaw, n8n), content generation can be completed in no network environment
4. Quantitative selection: --quant Q4_K_M parameters help you make a trade-off between speed and accuracy

Installation and use

AI Agent frameworks such as ```bash

pipx install whichllm

whichllm

RTX 5090

whichllm --gpu "RTX 5090"

whichllm run

Python

whichllm snippet "qwen 7b"

##
- [GitHub: WhichLLM](https://github.com/Andyyyy64/whichllm)
- [Show HN](https://news.ycombinator.com/item?id=48063953)

##

WhichLLM `llama.cpp`（GGUF ），`transformers`（AWQ/GPTQ ），`nvidia-ml-py`（GPU ）。 LLM ，`GGUF` 。`OpenClaw` are also often used in conjunction with local models.

## Internal link guidance
- Want to learn local models? Watch: [How to run local AI models on M4 Mac with LM Studio: A complete 30-minute tutorial](https://waytoclawearn.com/tutorials/guide-lm-studio-m4-local-ai-model-tutorial)
- Real case: [He Built an AI Automation Stack with Claude + n8n — $4K to $12K/mo in 6 Months](https://waytoclawearn.com/cases/agency-owner-claude-n8n-ai-automation-12000-month)
- DeepClaude cost savings: [Claude Code + DeepSeek V4 Building Tutorial: API Fees Dropped by 90% (15 Minutes)](https://waytoclawearn.com/tutorials/deepclaude-deepseek-claude-code-tutorial)