llama.cpp on the NVIDIA RTX 3090
On the NVIDIA RTX 3090, llama.cpp runs Qwen3.8 27B at 4-bit at a median 35.7 tokens per second for generation, from 2 plain runs (128K context, 1 not stated).
Speed by model on the NVIDIA RTX 3090
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 | 35.7 t/smedian of 2, 128K context, 1 not stated | vLLM 38.4 t/sone report, 144K to 200K context |
| Gemma 4 31B 4-bit | 36.5 t/sone report, context not stated | no other engine reported |
| Llama 3 8B 4-bit | 111.7 t/sone report, 1K context | no other engine reported |
| Muse 30B 4-bit | 30.0 t/sone report, 8K context | no other engine reported |
| Qwen2.5 32B 4-bit | 28.0 t/sone report, 32K context | no other engine reported |
| Qwen3.6 27B 4-bit | 42.0 t/sone report, 100K context | no other engine reported |
| Qwen3.6 35B · 3B active 3-bit | 120.0 t/sone report, 256K context | no other engine reported |
| Qwen3.6 35B · 3B active 4-bit | 101.7 t/sone report, 64K context | no other engine reported |
| Qwen3.8 27B 3-bit | 42.3 t/sone report, 16K context | no other engine reported |
61 llama.cpp reports on the NVIDIA RTX 3090
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 · Q4_K_XL66.4 t/s(speculative decoding)
- Qwen3.8 27B70.0 t/s(speculative decoding)
- GLM-5.3 320B · 18B active · GSQ-RCO 3.0bit Q4_K attention21.5 t/s(2 cards, part in system RAM)
- GLM-5.3 320B · 18B active · Q2_K Medium19.2 t/s(2 cards, speculative decoding, part in system RAM)
- Llama 3 8B · Q4_K_M111.7 t/s
- Step-3.5-Flash 196B · 11B active · IQ4_XS17.5 t/s(3 cards, part in system RAM)
- Llama 3 8B · Q4_0129.1 t/s(2 cards)
- Qwen3.8 27B · Q8_K_XL44.2 t/s(2 cards)
- Qwen3-Next 80B · 3B active · Q4_K_M94.5 t/s(part in system RAM)
- Qwen3.8 27B · Q8_K_XL48.8 t/s(2 cards, speculative decoding)
- Gemma 4 31B · Q4_K_S36.5 t/s
- Qwen3.8 27B38.1 t/s(2 cards)
- Ornith1.5 35B · 3B active · Q4_K_M169.4 t/s(speculative decoding)
- DeepSeek V4 Flash 284B · 13B active · IQ2_XXS24.5 t/s(4 cards, speculative decoding, part in system RAM)
- DeepSeek V4 Flash 284B · 13B active · UD-Q8_K_XL1.8 t/s(2 cards, part in system RAM)
- DeepSeek V4 Flash 284B · 13B active6.0 t/s(4 cards, part in system RAM)
- Muse 30B · Q4_K_M30.0 t/s
- Qwen3.8 27B · Q4_K_M36.4 t/s
- Qwen3.6 35B · 3B active · Q4_K_XL101.7 t/s
- Qwen3.8 27B · UD-Q4_K_XLPP 1080.0 t/s(2 requests at once)
- Qwen3.6 35B · 3B active · Q3_K_M120.0 t/s
- Qwen3.8 125B · 6B active · IQ4_XS40.0 t/s(part in system RAM)
- Qwen3.6 27B · Q4_K_M50.3 t/s(speculative decoding)
- Qwen3.8 27B · Q2_64140.0 t/s
- Qwen3.8 27B · 3.47 bpw42.3 t/s
- Qwen3.8 27B · IQ4_XS70.0 t/s(4 cards, speculative decoding)
- Qwen3.8 27B · IQ4_XS-M90.0 t/s
- Qwen3.8 125B · 6B active · IQ4_XS38.0 t/s(2 cards, part in system RAM)
- Qwen3.8 125B · 6B active · UD-Q4_K_XL49.0 t/s(2 cards, speculative decoding, part in system RAM)
- Qwen3.8 125B · 6B active · UD-Q4_K_XL34.3 t/s(2 cards, speculative decoding, part in system RAM)
Frequently asked
How fast is llama.cpp on the NVIDIA RTX 3090?
On the NVIDIA RTX 3090, llama.cpp runs Qwen3.8 27B at 4-bit at a median 35.7 tokens per second for generation, from 2 plain runs (128K context, 1 not stated).
Is llama.cpp faster than other engines on the NVIDIA RTX 3090?
Same card, same model size and bit class, plain runs only: Qwen3.8 27B at 4-bit gives llama.cpp 35.7 t/s (2 runs) and vLLM 38.4 t/s (1). Context length and engine build differ between the runs, so treat it as a lead, not a benchmark.