llama.cpp on the NVIDIA RTX 5090
On the NVIDIA RTX 5090, llama.cpp runs Qwen3.8 27B at 4-bit at a median 81.0 tokens per second for generation, from 2 plain runs (120K context, 1 not stated).
Speed by model on the NVIDIA RTX 5090
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 | 81.0 t/smedian of 2, 120K context, 1 not stated | no other engine reported |
| Gemma 4 26B · 4B active 4-bit | 243.9 t/sone report, context not stated | no other engine reported |
| LFM2.5 8B · 1B active 8-bit | 563.1 t/sone report, context not stated | no other engine reported |
| Qwen3 30B · 3B active 4-bit | 352.1 t/sone report, context not stated | no other engine reported |
| Qwen3.5 35B · 3B active 4-bit | 194.0 t/sone report, context not stated | no other engine reported |
| Qwen3.6 35B · 3B active 4-bit | 215.1 t/sone report, context not stated | no other engine reported |
| Qwen3.8 27B 6-bit | 58.1 t/sone report, 32K context | no other engine reported |
37 llama.cpp reports on the NVIDIA RTX 5090
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.
- Qwen3.8 27B · UD-Q4_K_XL701.2 t/s(speculative decoding)
- LFM2.5 8B · 1B active · Q8_0563.1 t/s
- Qwen3 30B · 3B active · Q4_K_M352.1 t/s
- Qwen3.8 27B · IQ4_XS94.0 t/s
- GLM-5.3 320B · 18B active · IQ4_XS22.0 t/s(2 cards, speculative decoding, part in system RAM)
- Qwen3.8 27B · NVFP4148.0 t/s(speculative decoding)
- Qwen3.8 27B · Q4_K_M68.0 t/s
- Qwen3.8 27B · Q6_K143.0 t/s
- Qwen3.8 125B · 6B active · UD-Q2_K_XL48.0 t/s(part in system RAM)
- Qwen3.8 27B · UD-Q6_K_XL98.0 t/s(speculative decoding)
- DeepSeek V4.1 Flash 552B · 16B active · MXFP45.1 t/s(part in system RAM)
- Qwen3.5 35B · 3B active · Q4_K_XL194.0 t/s
- Qwen3.8 125B · 6B active · UD-Q4_K_XL20.0 t/s(part in system RAM)
- Qwen3.8 125B · 6B active · Q4_K_M40.0 t/s(part in system RAM)
- Qwen3.8 27B · Q6_K97.9 t/s(speculative decoding)
- Qwen3.8 125B · 6B active · Q2PP 3000.0 t/s(2 cards)
- Muse 30B · UD-Q5_K_M253.0 t/s(speculative decoding)
- DeepSeek V4 Flash 284B · 13B active · mixed Q8/Q4/Q213.7 t/s(part in system RAM)
- DeepSeek V4 Flash 284B · 13B active · Q2_K21.3 t/s(part in system RAM)
- DeepSeek V4 Flash 284B · 13B active · Q8_K_XL17.0 t/s(part in system RAM)
- Qwen3.8 2400B · 95B active · UD-Q1_00.3 t/s(5 cards, part in system RAM)
- Qwen3.8 2400B · 95B active · Q1_00.8 t/s(2 cards, speculative decoding, part in system RAM)
- DeepSeek V4 Flash 284B · 13B active · Q8_K_XL18.0 t/s(part in system RAM)
- DeepSeek V4 Flash 284B · 13B active · Q8_K_XL30.2 t/s(2 cards, part in system RAM)
- Qwen3.8 27B · Q6_K58.1 t/s
- Qwen3.8 27B86.2 t/s(speculative decoding)
- Qwen3.8 27B · Q8_0100.0 t/s(speculative decoding)
- Qwen3.8 27B · Q5_K_L120.0 t/s(speculative decoding)
- Qwen3.8 27B · Q4_K_XL75.0 t/s(speculative decoding)
- Qwen3.8 27B · Q8_K_L97.7 t/s(speculative decoding)
Frequently asked
How fast is llama.cpp on the NVIDIA RTX 5090?
On the NVIDIA RTX 5090, llama.cpp runs Qwen3.8 27B at 4-bit at a median 81.0 tokens per second for generation, from 2 plain runs (120K context, 1 not stated).