marfrit / arcint

git clone https://github.com/marfrit/arcint.git

Intel Arc LLM inference server (A770/B60): Qwen hybrids up to Qwen3.8-Flash-Next (512-expert MoE through an expert cache in VRAM and host RAM) on llama.cpp with arcint's own OpenCL kernels, MTP speculation, context checkpoints; image generation, vision, speech in and out