Intel Arc B390 runs Gemma 4, Qwen 35B

Panther Lake's Xe3 architecture finally makes high-parameter local AI practical for thin-and-light hardware, though it remains a "prosumer" solution rather than a full NVIDIA replacement.

• –Throughput on Qwen 35B (MoE) hits 12-18 t/s, while denser models like Gemma 4 31B reach a solid 4-7 t/s via Intel's IPEX-LLM acceleration.
• –The 96 dedicated Xe3 Tensor Cores narrow the gap between integrated and discrete silicon, rivaling entry-level 45W dGPUs in matrix multiplication tasks.
• –Dynamic VRAM allocation allows the iGPU to utilize up to 90% of system RAM, enabling 32GB ultraportables to host large model weights that typically require 12GB+ of dedicated VRAM.
• –Software maturity is the lingering bottleneck; while Ollama and IPEX-LLM provide the foundation, the ecosystem still lacks the plug-and-play stability of NVIDIA's CUDA.
• –The 1.3kg MSI Prestige 14 AI+ effectively sets a new performance floor for "AI PCs," making local development of large models viable for mobile users.