In this briefing
- Acer announced the Veriton RI110 AI Mini Workstation, specifying 96GB of memory, dual network ports, a 120W power adaptor and future availability windows. ASUS showcased the ProArt GR1X mini PC but did not publish its price or launch date in the announcement.
- Both companies say their devices can support models with up to 120B parameters, yet the public materials do not specify a common model, quantisation precision, context length, concurrency level or token-throughput measure, so the figures do not support a like-for-like performance comparison.
- Qualcomm and ASUS plan to support more than 50 demonstration pharmacies in southern Taiwan with local devices. The release says the cloud-dependent GPT-OSS 120B deployment was adapted to a locally runnable 20B model. Deployment and outcome validation are still pending.
Device layer: compact systems have different delivery states
On 2 September, Acer announced the Veriton RI110 AI Mini Workstation. It uses an Intel Core Ultra X7 358H processor and Intel Arc B390 graphics, with configurations up to 96GB of dual-channel LPDDR5X memory and a 4TB PCIe Gen 4 SSD. Its dimensions are 138.5×131.3×52.1 mm and its approximate weight is 0.63 kg. The 120W figure is the power-adaptor rating; it does not directly establish sustained whole-system power.
The RI110 has one 10GbE network port, one 2.5GbE network port and Wi-Fi 7. A PCIe 4.0×4 OCuLink interface offers a connection of up to 64Gbps for an external GPU or NVMe storage. Acer says that the device supports models of up to 120B parameters, permits deployment of Agents including Qubi Claw, and provides an isolated sandbox. It is an announced product that has not yet shipped: availability is planned for North America in the fourth quarter of 2026 and for Europe, the Middle East and Africa in the first quarter of 2027; Acer has not published exact pricing or regional configurations.
At its IFA media showcase on the same day, ASUS announced the ProArt P16, P14 and GR1X powered by NVIDIA RTX Spark. The company lists up to 128GB of unified memory and up to 1 petaflop FP4 compute at peak. The GR1X measures 150×150×51 mm, includes 10GbE, Wi-Fi 7 and dual-fan cooling, and is described as a local Agent mini PC designed for sustained operation. The announcement did not provide prices, availability regions or launch dates for these ProArt devices, so they should currently be understood as announced and showcased rather than already available for purchase.
These two sets of updates give compact local-AI systems concrete product names, interfaces, memory specifications and delivery statuses beyond the prototype concept, but the products remain at different stages of maturity. For enterprise deployment, announcement, planned availability, a live showcase and stock availability are separate milestones. Only when final configurations, driver and model support, fleet management, warranty and delivery schedules are also defined do small devices provide an actionable procurement boundary.
Performance layer: ‘supports 120B’ is not a comparable operating result
The Acer and ASUS materials both cite a ceiling of 120B parameters, but they do not publish a common set of test conditions. Acer does not identify the model, numerical precision, context length, batch size or memory allocation involved. ASUS specifies peak FP4 compute and up to 128GB of unified memory, but likewise does not state the quantisation, context or concurrency setting for the 120B workload. Neither announcement provides time to first token, output tokens per second, sustained-load power or recovery data for a common model.
The software entry points also differ. Acer emphasises downloadable Agents, one-click installation of Qubi Claw, local workflows and an isolated sandbox. ASUS focuses on the CUDA ecosystem, ComfyUI, local FLUX and WAN workflows, and keeping prompts, files and creative assets on the device during offline operation. These are descriptions of each company's software scope; they do not mean that the two systems have passed the same security, compatibility or stability acceptance tests.
Large memory pools and low-precision compute put more models within the design space of desktop devices, but parameter count is only an initial filter. Until the model, precision, context, concurrency and output quality are aligned, 120B, 96GB, 128GB and peak FP4 compute do not form a performance ranking. Whether an enterprise Agent is suitable for permanent local operation also depends on isolation, identity and access, audit logs, model updates, remote recovery and long-duration stability, conditions that the public materials do not yet cover fully.
Field layer: the pharmacy programme is intended for frontline validation
On 2 September, Qualcomm and ASUS announced the launch of the Pharmaceutical AI Agent programme in southern Taiwan. The companies say that AI PC systems and edge devices carrying the Agent will be donated to support more than 50 demonstration pharmacies across Chiayi, Tainan, Kaohsiung and Pingtung. In the first phase, local health authorities will coordinate the participation of community pharmacies, with the aim of validating the feasibility, usability and effectiveness of AI-assisted medication review in real settings. The figure defines the planned demonstration scope; it does not mean that more than 50 pharmacies are already live.
The planned deployment includes a Zenbook A16 AI PC powered by Snapdragon X2 Elite Extreme and an Aetina MegaEdge AIP-FR68 on-premises edge device equipped with Qualcomm Cloud AI 100 Ultra. The announcement says that the open-weight GPT-OSS deployment was adapted from using a cloud-dependent 120B large model to a 20B model that can run locally on an AI PC, with access to the public drug package-insert database of the Taiwan Food and Drug Administration. The Agent can perform inference locally in offline environments, reducing the need to transmit sensitive medical and medication data to an external cloud.
This remains a demonstration and validation programme. The announcement does not disclose the model-optimisation method, comparative task accuracy, incorrect-suggestion rate, response latency, concurrency, device pricing or operational responsibility, and it provides no clinical outcomes. Connecting to public medicine information does not demonstrate regulatory authorisation as a medical device or clinical validation for medication decisions. Local operation also does not automatically show that access control, encryption, audit and medical-privacy compliance requirements have been met.
Compared with a hardware showcase alone, the pharmacy programme is intended to test local AI within specific processes for offline operation, sensitive-data handling and professional human review. Only if later reporting covers task quality, pharmacist review, version updates, audit and exception handling will it become possible to judge whether a local Agent can operate reliably in a frontline workflow. The current evidence confirms the planned scope, device combination and direction of model optimisation; it cannot establish that medication safety or operating efficiency has already improved.
What to watch next
- Final pricing, regional configurations, actual delivery dates, enterprise fleet-management capabilities and long-term support policies for the Acer RI110 and ASUS ProArt RTX Spark devices.
- The specific model, quantisation precision, context length, sustained power, time to first token, throughput, concurrency and output quality for the two types of 120B local workload.
- Human-review processes, error and refusal metrics, audit controls, privacy compliance and real outcomes for the Pharmaceutical AI Agent in demonstration pharmacies.
Sources and verification
Golden Data has edited this briefing from the public materials listed above. The original sources govern facts and figures. The enterprise relevance sections are Golden Data editorial analysis and do not constitute an endorsement of any third-party product.
← Back to AI Daily Briefing