BYRAM AI LAB · 2× INTEL ARC PRO B70 32 GB · 64 GB DDR5
Local AI, measured
I run open models on Intel Arc GPUs for real work: local agents, coding, image generation. I measure everything that runs, and write down what happens. Every run ships with the config, so you can reproduce it.
6
RUNS LOGGED
105.8 t/s
BEST WARM DECODE
3
ENGINES COMPARED
WARM DECODE · TOK/S · 6 RUNS
Qwen3.6-35B-A3B · vLLM XPU · ASRock B70105.8
Qwen3.8-27B · exl3xpu · ASRock B7057.0
Qwen3.8-27B · llama.cpp SYCL/XPU · ASRock B7056.8
Qwen3.8-27B · exl3xpu · ASRock B7056.6
Qwen3.8-27B · llama.cpp SYCL/XPU · ASRock B7056.1
Qwen3.8-27B · llama.cpp SYCL/XPU · ASRock B7044.9
LATEST RUNS · 1× INTEL ARC PRO B70
Qwen3.6-35B-A3B · GPTQ-Int4 · vLLM XPU
105.8 t/s warm
TTFT 1.50 s · VRAM 31.4 GB peak
2026-10-16baseline
Qwen3.8-27B · EXL3 4.00bpw · exl3xpu
57.0 t/s warm
TTFT 1.43 s · VRAM 30.7 GB peak
2026-10-16baseline
Qwen3.8-27B · EXL3 · batch 4096→8192
56.6 t/s warm
TTFT 1.40 s · VRAM 31.4 GB peak
2026-10-17▼ −0.4 vs prev
Qwen3.8-27B · GGUF IQ3_S · llama.cpp
44.9 t/s warm
TTFT 23.45 s · VRAM 20.9 GB peak
2026-10-17baseline
Qwen3.8-27B · GGUF · ubatch 4096 + KV q8_0
56.8 t/s warm
TTFT 16.00 s · VRAM 29.8 GB peak
2026-10-17▲ +11.9 vs prev
Qwen3.8-27B · GGUF · KV q8_0
56.1 t/s warm
TTFT 16.65 s · VRAM 23.4 GB peak
2026-10-17▼ −0.7 vs prev