Acer Veriton RI110 Runs 120B AI Models in 0.63kg

A 0.63kg Box Aimed at Local Model Inference

Acer introduced the Veriton RI110 on September 2, 2026, and the spec sheet reads like a deliberate answer to a question a lot of people have been asking: what does a mini PC look like when local inference is the design target rather than a bullet point? The answer here is a 138.5 x 131.3 x 52.1 mm chassis weighing roughly 0.63 kg, built around Intel's Core Ultra X7 358H with Arc B390 graphics, and fed by up to 96 GB of dual-channel LPDDR5X.

• Silicon: Intel Core Ultra X7 358H (Panther Lake) with Intel Arc B390 graphics, qualifying as a Copilot+ PC
• Memory and storage: up to 96 GB dual-channel LPDDR5X and up to 4 TB of M.2 2280 PCIe Gen 4 SSD
• Expansion: OCuLink over a dedicated PCIe 4.0 x4 link, rated at up to 64 Gbps for external GPUs or NVMe
• Availability: North America in Q4 2026, EMEA in Q1 2027; Acer has not published pricing

Why 96GB of LPDDR5X Is the Headline Number

Acer says the RI110 supports AI models of up to 120 billion parameters. That is the vendor's own claim and it depends heavily on quantisation — a 120B model does not fit in 96 GB at anything like full precision, so this is a statement about aggressively quantised weights rather than a general guarantee. Taken on those terms it is still meaningful, because the constraint on running large models locally has almost always been unified memory capacity rather than raw compute.

That is the same pattern we saw when the Mac mini M6 cut local LLM prompt latency by 4.8x — the win came from memory bandwidth and capacity, not from a faster core count. A Panther Lake box with 96 GB on a dual-channel LPDDR5X bus is playing the same game from the x86 side.

What Does OCuLink Add to a Mini Workstation?

The RI110 exposes OCuLink on a dedicated PCIe 4.0 x4 interface, good for up to 64 Gbps. In practice that means an external GPU enclosure or a fast external NVMe array attaches at close to native PCIe speed rather than through a Thunderbolt bottleneck.

For a single board this small, that changes the shape of the upgrade path. You start with the integrated Arc B390 for light inference and content work, and when a workload outgrows it you hang a discrete GPU off the back without replacing the machine. OCuLink on compact systems has been quietly spreading — the GPD Box paired Panther Lake with a PCIe 5.0 eGPU port earlier this year — and it is becoming the differentiator between a mini PC and a mini workstation.

The Rest of the I/O

Networking is dual LAN, 10GbE alongside 2.5GbE, plus Wi-Fi 7 and Bluetooth 5.4. Display and peripheral duty falls to two rear USB4 Type-C ports, additional USB 3.2 and USB 2.0, HDMI 2.1 and DisplayPort. There is a Kensington lock slot, which sounds mundane until you remember these boxes end up bolted under desks in shared offices and lab benches.

Acer also lists sandboxed environments for AI deployment and compatibility with its Qubi Claw agentic AI software, and ships Windows 11 Pro or Home depending on configuration.

How Does This Compare to Acer's Last AI Mini Workstation?

The RI110 is the Intel counterpart to the Veriton RA110, the six-inch Ryzen AI Max+ 395 mini workstation Acer showed in May. Same product family, same design brief, different silicon vendor — and notably, the Intel version brings OCuLink and dual LAN that the AMD box did not lead with.

The open question is price, and Acer has not answered it. Availability is Q4 2026 in North America and Q1 2027 in EMEA, with configurations varying by region. Until a number lands, the RI110 is a promising spec sheet rather than a recommendation. Our mini PC coverage will pick it up again when pricing is public.

Sources: Acer press release via PR Newswire — September 2, 2026; Liliputing — September 2, 2026.