Recommended models for RTX 4090 24GB

11 of 13 models fit fully in GPU memory at 16k context. Fastest three:

Qwen3-Coder 30B-A3B Q4_K_M	~266 tok/s measured
gpt-oss 20B Q8_0	~258 tok/s measured
Qwen3.6 35B-A3B UD-Q4_K_M	~190 tok/s measured

All models and commands · Data: Local LLM Advisor (localllmadvisor.com), CC BY 4.0