Intel LLM Scaler adds Multi-token Prediction, LoRA support for Arc Pro GPUs

LLM Scaler represents Intel's strategy to enable its Arc Pro GPUs as viable alternatives for generative AI workloads. Rather than building proprietary software, the platform leverages open-source frameworks (vLLM, ComfyUI, SGLang Diffusion, Xinference) to ensure compatibility and ease of adoption for developers already familiar with these tools. The August 2026 release adds Multi-token Prediction and LoRA Serving—two features that improve both inference speed and fine-tuning capabilities—extending the platform's utility for both deployment and model customization.

The breadth of supported models has grown substantially: the platform now covers major open-weight models from OpenAI (GPT-OSS), Qwen, DeepSeek, Llama, Mistral, Gemma, and others, spanning text, vision, embedding, and omni (multimodal) architectures. Quantization options (INT4, FP8, MXFP4) allow operators to trade off precision for memory and speed, a critical capability for running larger models on GPU memory-constrained systems. The Omni Serving mode, which exposes image and audio generation via OpenAI-compatible APIs, lowers integration friction for applications already built on that interface.