The inference runtime Arc deserves
~31 tok/s on Arc Pro B50. Context window adjustable per model —
ships with 32K by default. 4 concurrent sessions, confirmed coherent.
Free for personal use.
$
curl -fsSL https://gsisystems.net/stratum/install.sh | bash
Supported Hardware
Arc Pro B70
32 GB
flagship AI workstation
Arc Pro B60
24 GB
professional multi-model
Arc Pro B50
16 GB
production tested
Arc B580
12 GB
consumer sweet spot
Arc B570
10 GB
entry Battlemage
Requirements
- Linux x86_64
- Intel Arc Xe2 GPU (Battlemage architecture)
- Linux kernel
6.17+recommended (required for B70) - Use Intel Xe drivers — not Vulkan
Why STRATUM
Standard llama.cpp on Arc
2.7
tok/s
STRATUM on Arc
~31
tok/s · ~11×
Same hardware. Same model. ~11× difference.
Models
STRATUM runs any GGUF model from HuggingFace.
stratum run qwen2.5-coder-7b
stratum run deepseek-r1-7b
stratum run llama3.2-3b
Free for personal use. Commercial license required for production deployments.
[email protected]
[email protected]