ik_llama.cpp benchmarks
An inference engine for running open-weight LLMs locally.
12 community reports
ik_llama.cpp results by GPU
Hardware with reported ik_llama.cpp results, with each report's model, quant and speed, newest first. Runs on several cards, with speculative decoding, with batched requests or with part of the model in system RAM say so, since each describes a different setup. Prompt processing is how fast the input is read, not how long you wait for the first word. How to measure time to first token.
ik_llama.cpp on NVIDIA RTX 30902 reports
- Qwen3.6 27B · IQ4_KS72.0 t/s(speculative)
- Qwen3.6 27B · IQ4_KS72.9 t/s, 1,261 t/s prompt(speculative)
ik_llama.cpp on NVIDIA RTX 5070 Ti2 reports
- Raon-1 29B · 4B active · IQ3_KS232.0 t/s(part in system RAM)
- Gemma 4 26B · 4B active · Q4_K_XL6.9 t/s, 2,181 t/s prompt(part in system RAM)
ik_llama.cpp on NVIDIA RTX Pro 6000 Blackwell2 reports
- GLM-4.5-Air 106B · 12B active · UD-Q3_K_XL8.3 t/s, 444 t/s prompt(part in system RAM)
- Qwen3.8 125B · 6B active · IQ440.0 t/s, 2,000 t/s prompt
ik_llama.cpp on NVIDIA RTX 3060 12GB1 report
- Qwen3.8 125B · 6B active · Q4_K_XL22.1 t/s, 66 t/s prompt(speculative, part in system RAM)
ik_llama.cpp on NVIDIA RTX 40701 report
- Qwen3.6 35B · 3B active · Q3_K_P28.7 t/s(part in system RAM)
ik_llama.cpp on NVIDIA RTX 4070 Super1 report
- Qwen3.6 35B · 3B active · IQ4_XS-4.19bpw110.2 t/s(speculative, part in system RAM)
ik_llama.cpp on NVIDIA RTX 50901 report
- Qwen3.8 125B · 6B active · AP Q4_K_M38.8 t/s, 200 t/s prompt(part in system RAM)
ik_llama.cpp on NVIDIA T4 16GB1 report
- Qwen3.8 125B · 6B active · UD-Q4_K_XL17.6 t/s, 160 t/s prompt(part in system RAM)
ik_llama.cpp on NVIDIA V100 16GB1 report
- Qwen3.5 9B · Q6_K4.6 t/s, 53 t/s prompt
ik_llama.cpp, GPU not identified8 reports
- Raon-1 3.1B · Q4_K_M23.5 t/s, 229 t/s prompt
- DeepSeek V4.1 Flash 552B · 16B active · Q6_K9.9 t/s, 101 t/s prompt(part in system RAM)
- Bonsai 2 4B · Q2_011.6 t/s, 49 t/s prompt
- Bonsai 1.7B · PQ2_071 t/s prompt
- Qwen3.6 35B · 3B active · Q4_K_M23.0 t/s(part in system RAM)
This engine doesn't yet have an editorial profile on llamaperf. The community reports below show how it's been used in practice across different hardware.
Compare GPUs running ik_llama.cpp
Every card is compared on one model, Qwen3.8 125B · 6B active at 4-bit, the size and bit class that most cards here have plain runs of. A card with no run of that model shows none rather than a figure from a different model.
| GPU | VRAM | Reports | Median speed, Qwen3.8 125B · 6B active 4-bit |
|---|---|---|---|
| NVIDIA RTX 3090nvidia | 24GB | 3 | no Qwen3.8 125B · 6B active 4-bit run reported |
| NVIDIA RTX 5070 Tinvidia | 16GB | 2 | no Qwen3.8 125B · 6B active 4-bit run reported |
| NVIDIA RTX Pro 6000 Blackwellnvidia | 96GB | 2 | Gen 40.0 t/sPP 2000.0 t/s · n=1 |
| NVIDIA RTX 3060 12GBnvidia | 12GB | 1 | no Qwen3.8 125B · 6B active 4-bit run reported |
| NVIDIA RTX 4070nvidia | 12GB | 1 | no Qwen3.8 125B · 6B active 4-bit run reported |
| NVIDIA RTX 4070 Supernvidia | 12GB | 1 | no Qwen3.8 125B · 6B active 4-bit run reported |
| NVIDIA RTX 5090nvidia | 32GB | 1 | no Qwen3.8 125B · 6B active 4-bit run reported |
| NVIDIA T4 16GBnvidia | 16GB | 1 | no Qwen3.8 125B · 6B active 4-bit run reported |
ik_llama.cpp against other engines
Pairs of plain runs on the same card, of the same model size at the same bit class: one device, one request, no speculative decoding, the whole model in memory. Context length and build still differ between the two sides, and each side shows its own.
| Model and GPU | ik_llama.cpp | Other engine |
|---|---|---|
| Qwen3.8 125B · 6B active 4-bitNVIDIA RTX Pro 6000 Blackwell | 40.0 t/sPrompt processing (prefill) 2000.0 tokens/s (1)one generation report, 200K context | llama.cpp105.5 t/sPrompt processing (prefill) 1955.0 tokens/s (1)median of 2 generation reports, 2K to 240K context |
| Qwen3.8 125B · 6B active 4-bitNVIDIA RTX Pro 6000 Blackwell | 40.0 t/sPrompt processing (prefill) 2000.0 tokens/s (1)one generation report, 200K context | NInfer172.0 t/sPrompt processing (prefill) 5905.0 tokens/s (1)one generation report, 512 context |
| Qwen3.8 125B · 6B active 4-bitNVIDIA RTX Pro 6000 Blackwell | 40.0 t/sPrompt processing (prefill) 2000.0 tokens/s (1)one generation report, 200K context | SGLang-Prompt processing (prefill) 11352.0 tokens/s (1)no generation reports, 256K context |
Top models on ik_llama.cpp
Frequently asked
Is ik_llama.cpp faster than other engines?
llamaperf has 2 matched pairs of ik_llama.cpp and other-engine runs on the same card, model size and bit class, but none is comparable enough to say which engine is faster: the sides have too few reports or were run at different context lengths. A pair counts toward a direction only when each side has at least 3 generation reports with a stated context, and the two sides' contexts overlap. The measured figures and their conditions are in the table on this page.