ASUS Korea (hereinafter referred to as ASUS) announced that it has launched the UGen Intel Arc Pro B70 32GB graphics card for local agentic AI environments. It features 32GB of ECC-based GDDR6 memory and can scale up to 128GB in dual or quad configurations.
According to ASUS, a single card supports most models with 27 billion parameters or fewer. When running Llama 3.1 8B, it supports a context window of 128K tokens, and considering the development environments of small-to-medium enterprises and research labs, it offers multi-card configurations.
Software compatibility includes Intel LLM Scaler, vLLM XPU, llama.cpp (SYCL and Vulkan), Ollama, Intel OpenVINO, oneAPI, and PyTorch XPU. It also supports ComfyUI for image generation. ASUS explained that it has verified model operation for Llama 3.1/3.2 (8B–70B), Qwen 3.x, DeepSeek-R1 Distill (7B–70B), Mistral, Mixtral, OpenAI gpt-oss (20B, 120B), Whisper-large-v3, InternVL3.5, and more.
This product targets desktop environments processing large language models (LLMs) and multi-agent AI workloads. ASUS stated that it configured the memory capacity and architecture to handle the increased KV-cache usage from long-context reasoning as well as the memory overhead of multi-agent deployments. Additionally, ASUS highlighted the ability to build AI deployment environments while keeping sensitive data securely stored in local environments.
ASUS introduced the product as a solution for its agent computer vision spanning cloud, local, and hybrid AI deployment environments. Detailed information about the product and agent computer vision can be found on the official ASUS website.
