Benchmarks

Every test we have run, with the exact conditions. Numbers are generated from the raw llama-bench output, which you can download.

4 test sessions. Last run Sep 26, 2026.

T0004 AMD Radeon Graphics, integrated (Ryzen 3 4300U)

Machine:: ACEMAGIC K1
Date:: Sep 26, 2026
Backend:: Vulkan
Software:: llama.cpp build 11191 (4b1a27fa0)
System:: AMD Ryzen 3 4300U (4 cores / 4 threads, Zen 2); 16 GB DDR4-2666 (one SO-DIMM as shipped, so single channel; the second slot is free)

Run over a Remote Desktop session. The K1 has one 16 GB DDR4-2666 module, so its memory runs single-channel at 21.3 GB/s (theoretical peak). Windows power mode: Best performance. The integrated graphics share that memory. gpt-oss 20B (12.1 GB) did not fit in the memory Windows lets the integrated graphics use, and failed to load.

T0004: AMD Radeon Graphics, integrated (Ryzen 3 4300U)
Model	Reading the prompttokens/s	Writing the answertokens/s
Gemma 4 E4B Q4_0 Whole model on GPU	54.1	5.87
Gemma 4 E4B Q4_0 Whole model on GPU, with 4,096 tokens already in context	9.38	5.08
gpt-oss 20B MXFP4 Whole model on GPU	failed	failed
Llama 2 7B Q4_0 Whole model on GPU	38.9	4.82
Llama 2 7B Q4_0 Whole model on GPU, with 4,096 tokens already in context	17.2	2.79
Qwen3.5 9B Q4_K_M Whole model on GPU	33.6	3.37
Qwen3.5 9B Q4_K_M Whole model on GPU, with 4,096 tokens already in context	14.8	3.10

T0004: AMD Radeon Graphics, integrated (Ryzen 3 4300U). Test T0004 on our acemagic k1, AMD Radeon Graphics, integrated (Ryzen 3 4300U). Measured with llama.cpp's llama-bench (build 11191 (4b1a27fa0)): prompt speed over 512 tokens, answer speed over 128 tokens, average of 5 runs. Higher is better.

T0003 AMD Ryzen 3 4300U (CPU only)

Machine:: ACEMAGIC K1
Date:: Sep 26, 2026
Backend:: CPU
Software:: llama.cpp build 11191 (4b1a27fa0)
System:: AMD Ryzen 3 4300U (4 cores / 4 threads, Zen 2); 16 GB DDR4-2666 (one SO-DIMM as shipped, so single channel; the second slot is free)

Run over a Remote Desktop session. The K1 has one 16 GB DDR4-2666 module, so its memory runs single-channel at 21.3 GB/s (theoretical peak). Windows power mode: Best performance.

T0003: AMD Ryzen 3 4300U (CPU only)
Model	Reading the prompttokens/s	Writing the answertokens/s
Gemma 4 E4B Q4_0 CPU only	31.8	4.79
gpt-oss 20B MXFP4 CPU only	27.4	5.50
Llama 2 7B Q4_0 CPU only	20.5	3.94
Qwen3.5 9B Q4_K_M CPU only	18.1	2.70

T0003: AMD Ryzen 3 4300U (CPU only). Test T0003 on our acemagic k1, AMD Ryzen 3 4300U (CPU only). Measured with llama.cpp's llama-bench (build 11191 (4b1a27fa0)): prompt speed over 512 tokens, answer speed over 128 tokens, average of 5 runs. Higher is better.

T0002 NVIDIA GeForce RTX 4060 8 GB

Machine:: Reference desktop
Date:: Sep 25, 2026
Backend:: CUDA 12.4, driver NVIDIA 610.88
Software:: llama.cpp build 11191 (4b1a27fa0)
System:: AMD Ryzen 9 7900X (12 cores / 24 threads, Zen 4); 32 GB DDR5-6400 (2 x 16 GB, dual channel)

In this machine the RTX 4060 runs on a PCIe 3.0 x1 link (the card supports PCIe 4.0 x8). Rows where the model is split between the GPU and system RAM (gpt-oss 20B) are slowed by that link, prompt reading most of all; rows with the whole model on the GPU are not affected.

T0002: NVIDIA GeForce RTX 4060 8 GB
Model	Reading the prompttokens/s	Writing the answertokens/s
Gemma 4 E4B Q4_0 Whole model on GPU	2,964	67.7
Gemma 4 E4B Q4_0 Whole model on GPU, with 4,096 tokens already in context	2,722	65.1
gpt-oss 20B MXFP4 GPU, expert weights in system RAM	51.7	25.2
gpt-oss 20B MXFP4 Split automatically between GPU and system RAM	106	39.5
Llama 2 7B Q4_0 Whole model on GPU	2,716	65.5
Llama 2 7B Q4_0 Whole model on GPU, with 4,096 tokens already in context	1,874	43.3
Qwen3.5 9B Q4_K_M Whole model on GPU	1,897	45.5
Qwen3.5 9B Q4_K_M Whole model on GPU, with 4,096 tokens already in context	1,808	44.6

T0002: NVIDIA GeForce RTX 4060 8 GB. Test T0002 on our reference desktop, NVIDIA GeForce RTX 4060 8 GB. Measured with llama.cpp's llama-bench (build 11191 (4b1a27fa0)): prompt speed over 512 tokens, answer speed over 128 tokens, average of 5 runs. Higher is better.

T0001 AMD Ryzen 9 7900X (CPU only)

Machine:: Reference desktop
Date:: Sep 25, 2026
Backend:: CPU
Software:: llama.cpp build 11191 (4b1a27fa0)
System:: AMD Ryzen 9 7900X (12 cores / 24 threads, Zen 4); 32 GB DDR5-6400 (2 x 16 GB, dual channel)

T0001: AMD Ryzen 9 7900X (CPU only)
Model	Reading the prompttokens/s	Writing the answertokens/s
Gemma 4 E4B Q4_0 CPU only	198	17.8
gpt-oss 20B MXFP4 CPU only	109	21.8
Llama 2 7B Q4_0 CPU only	134	15.5
Qwen3.5 9B Q4_K_M CPU only	92.5	10.4

T0001: AMD Ryzen 9 7900X (CPU only). Test T0001 on our reference desktop, AMD Ryzen 9 7900X (CPU only). Measured with llama.cpp's llama-bench (build 11191 (4b1a27fa0)): prompt speed over 512 tokens, answer speed over 128 tokens, average of 5 runs. Higher is better.