Machine Updated 20 October 2026

RTX 4090 desktop

My desktop with an NVIDIA GeForce RTX 4090: 24GB of graphics memory, a Ryzen 9 7950X and 96GB of memory.

NVIDIA's GeForce RTX 4090 graphics card — NVIDIA's GeForce RTX 4090. Image: NVIDIA

Measured on mine measured by me

Set-up	Model	Writing code	Writing prose	Reading (prefill)	Tested
RTX 4090 desktopllama.cpp on CUDA	Gemma 4 12BGoogle's own 4-bit QAT file (Q4_0), GGUF	99.7 tok/s	101.9 tok/s	5,072 tok/s	20 Oct 2026
RTX 4090 desktopllama.cpp on CUDA	Gemma 4 26B-A4BUD-IQ4_XS, about 4 bits, GGUF	161.4 tok/s	165.4 tok/s	7,257 tok/s	20 Oct 2026
RTX 4090 desktopllama.cpp on CUDA	gpt-oss-20bMXFP4, GGUF, thinking at its lowest level	238.6 tok/s	241.1 tok/s	9,940 tok/s	20 Oct 2026
RTX 4090 desktopllama.cpp on CUDA	Qwen 3.5 9BQ4_K_M, GGUF	137.6 tok/s	138.3 tok/s	8,584 tok/s	20 Oct 2026
RTX 4090 desktopLM Studio, llama.cpp on CUDA	Qwen 3.8 27BQ4_K_M, GGUF	62.4 tok/s	45.0 tok/s	1,994 tok/s	11 Oct 2026
RTX 4090 desktopPrism ML's llama.cpp on CUDA	Ternary Bonsai 2 27BPTQ1_0, about 1.75 bits a weight, GGUF	105.5 tok/s	106.1 tok/s	2,941 tok/s	20 Oct 2026

One user, the model's thinking switched off, the same prompts and the same test each time. All of my results, and how I test, are on the hardware leaderboard.

The hardware, exactly

Graphics card: NVIDIA GeForce RTX 4090 with 24GB of its own memory.
Processor: AMD Ryzen 9 7950X, 16 cores and 32 threads.
Memory: 96GB of DDR5 as two 48GB modules running at 6,000MT/s (Corsair CMK96GX5M2B6000Z30).
Motherboard: ASRock X670E Taichi.
System: Windows 11 Pro.