llama.cpp on the NVIDIA RTX 4090
On the NVIDIA RTX 4090, llama.cpp runs Qwen3.8 27B at 4-bit at a median 31.5 tokens per second for generation, from 2 plain runs (128K to 194K context).
Speed by model on the NVIDIA RTX 4090
Plain runs only: one device, one request, no speculative decoding, the whole model in memory. Each row is one model size at one bit class, with the other engines' plain runs of the same model on this card beside it. Context length and build still differ between runs, and each side shows its own.
| Model | llama.cpp | Other engines |
|---|---|---|
| Qwen3.8 27B 4-bit | 31.5 t/smedian of 2, 128K to 194K context | no other engine reported |
| Gemma 4 26B · 4B active 5-bit | 129.0 t/sone report, 256K context | no other engine reported |
| Qwen2.5 7B 4-bit | 158.9 t/sone report, context not stated | no other engine reported |
14 llama.cpp reports on the NVIDIA RTX 4090
Newest first. A run on several cards, with speculative decoding, with part of the model in system RAM or with batched requests says so, since each describes a different setup.
- Qwen3.5 122B · 10B active · UD-IQ4_XS18.8 t/s(2 cards, part in system RAM)
- Qwen3.8 27B · Q4_K_M80.0 t/s
- Qwen2.5 7B · Q4_K_M158.9 t/s
- Qwen3.8165.9 t/s(speculative decoding)
- Qwen3.8 27B · IQ4_XS115.0 t/s(speculative decoding)
- Model not identified · Q480.0 t/s(speculative decoding)
- Gemma 4 26B · 4B active · Q5_K_M129.0 t/s
- Qwen3.8 27B · Q4_K_XL23.0 t/s
- DeepSeek V4 Flash 284B · 13B active · IQ4_XS3.0 t/s(2 cards)
- Qwen3.8 27B · Q4_K_M40.0 t/s
- Qwen3.8 27B · Q4_K_XL134.0 t/s(speculative decoding)
- Qwen3.8 27B · IQ4_XS90.0 t/s(speculative decoding)
- DeepSeek V4 Flash 284B · 13B active · UD-Q2_K_XL10.9 t/s(part in system RAM)
- Qwen3.6 27B · Q4_K_M80.0 t/s(speculative decoding)
Frequently asked
How fast is llama.cpp on the NVIDIA RTX 4090?
On the NVIDIA RTX 4090, llama.cpp runs Qwen3.8 27B at 4-bit at a median 31.5 tokens per second for generation, from 2 plain runs (128K to 194K context).