HP unveils ZBook Ultra G3a laptop built to run large AI models on-device

HP Inc. announced the HP ZBook Ultra G3a 16”, a thin-and-light mobile workstation pitched as a “personal AI factory” for professionals. The company says the AMD-powered system can run large language models with up to about 300 billion parameters locally, comes with Perplexity Computer preloaded (including Autodesk Revit support), and is expected to be available in October 2026. Pricing will come closer to launch (company).

HARDWARE desk — same-day company primary that a major PC vendor put a shipping-window mobile workstation on the table for large local LLMs and agentic workflows inside real design apps, not another vague “AI PC” slogan.

Local AI, as the same newsroom has it: the system is powered by up to an AMD Ryzen AI Max+ PRO 495 and up to 192GB of unified memory, with up to 160GB dedicated as VRAM, plus a 55 TOPS NPU and AMD Radeon integrated graphics. Unified memory means one pool of RAM shared by the processor and the graphics block. VRAM is the slice of that pool used as graphics memory — the space that holds model weights when a large language model, or LLM, runs on the machine. An NPU is a neural processing unit, a chip block built for AI math. TOPS means trillions of operations per second, a vendor yardstick for that block. HP says the laptop can run LLMs with up to about 300 billion parameters locally and can run multi-agent work alongside professional apps at the same time. On-device, or local, AI means the model runs on the laptop, not only in the cloud. Agentic AI here means software that can take multi-step actions with tools, not just answer one chat turn.

Software, still company: Perplexity Computer comes preloaded. HP says professionals can use local, on-device AI with Autodesk Revit through Perplexity Portable Computer for Windows’s MCP Connector — keep project data on the laptop, and optionally call a cloud model when a step needs it. MCP is the Model Context Protocol, a way an AI agent connects to an app such as Revit. HP and Perplexity also say they co-developed task-based Skills sold only to HP customers, meant to start work from a defined job rather than an empty prompt box.

Form factor, company: the 16-inch machine starts at about 4.2 pounds and 17.9 mm thin. A new thermal design claims up to about 100W of performance and up to 81.8% higher TDP versus the prior ZBook Ultra G1a, which the same footnote pegs at a 55W max. TDP is thermal design power — the heat and power budget the chip is allowed to sustain. File 4.2 pounds, 17.9 mm, 100W, and 81.8% as HP’s, with that G1a footnote.

Named workloads on the same page, still company: Autodesk Revit 2027, Autodesk 3ds Max, and Chaos Enscape for design and visualization, plus local AI agents. Extra newsroom color — HP Go 5G connectivity and an IT-replaceable keyboard — stays in Sources.

Portfolio context, company: HP presents the ZBook Ultra G3a as a new anchor next to HP ZGX Nano, HP ZGX Fury, HP Boost, and HP Remote System Controller. It also points to a Local AI Value Calculator on HP.com, pitched so Z workstation customers can estimate what moving some AI work from the cloud onto a local machine might save.

Availability, company and forward-looking: the HP ZBook Ultra G3a 16” is expected in October 2026. Pricing will be provided closer to availability. None was printed.

Named voices on the newsroom: Jim Nottingham, senior vice president and division president, Advanced Compute and Solutions, HP Inc.; Jack Huynh, senior vice president and general manager of the Computing and Graphics Group at AMD; and Ryan Foutty, vice president of business at Perplexity. Their quotes are color only and stay in Sources.

Plain English for the rest of the card: on-device / local AI = the model runs on the laptop, not only in the cloud. LLM = large language model. agentic AI = software that can take multi-step actions with tools. NPU = neural processing unit, a chip block for AI math. MCP = Model Context Protocol, how an AI agent connects to apps like Revit. VRAM = graphics memory used to hold model weights. TDP = thermal design power, the heat and power budget the chip is allowed. unified memory = one RAM pool shared by the processor and the graphics block.