Hermes Agent's Qwen3.8 27B local AI runs on Intel Arc Pro B70

GIGAZINE had previously tasked Hermes Agent with fetching and summarizing its own articles, using models such as Nous Research's Hermes 4 70B and Solar Pro 4. Solar Pro 4 eventually managed the job, but only after struggling: it failed to launch the browser, switched methods, and opened the wrong articles along the way. The premise of this latest test was that 32GB of VRAM would allow Qwen3.8 27B to run at a far higher-quality quantization setting than before.

The test did not go smoothly at first. Hermes Agent loaded the model into ordinary RAM instead of the GPU, and the runtime listed its engine as "llama.cpp b10964 (cpu)" — the previous (cuda) label was gone. Settings also turned out to be profile-specific, so the Vulkan backend had to be set for the correct profile and the gateway restarted. Even after that, the standard 27B model used nearly 30GB of VRAM with over 4GB spilling into RAM, leaving inference at single-digit tokens per second.

The outcome suggests that for readers running AI agents on a single high-VRAM GPU, the usable speed may depend less on raw memory than on picking a quantization that fits entirely in VRAM and getting the backend configuration right. The author still plans to test Bot Mode next, which may show whether the same setup holds up for agentic browsing tasks.