US startup Tiiny AI Inc. claims its pocket-sized AI supercomputer can run 120B-parameter open source models locally, eliminating cloud dependence and putting advanced AI directly into users’ hands.
US deep-tech startup Tiiny AI Inc. claims it has built the world’s smallest personal AI supercomputer, capable of running up to a full 120-billion-parameter large language model entirely on-device, without cloud connectivity, servers, or high-end GPUs.
The device, called Tiiny AI Pocket Lab, has been officially verified by Guinness World Records under the category “The Smallest MiniPC (100 LLM Locally)”, marking what the company positions as a first in AI supercomputing. The breakthrough signals a shift away from cloud-centric AI towards personal, portable, and private intelligence.
Designed to make advanced AI accessible to individual users—particularly in resource-constrained environments—the pocket-sized system reduces operational costs, latency, energy consumption, and privacy risks associated with cloud-based AI. The device measures 14.2 × 8 × 2.53 cm, weighs around 300 grams, and operates as a complete AI inference system.
Tiiny AI Pocket Lab is powered by an ARMv9.2 12-core CPU paired with a dedicated NPU delivering approximately 190 TOPS, supported by 80GB of LPDDR5X memory and 1TB of storage, all within a 65W power envelope. The company says it operates in the “golden zone” of personal AI, suitable for over 80 per cent of real-world use cases, with performance comparable to GPT-4o.
A core differentiator is its open-source foundation. The system uses TurboSparse, a neurone-level sparse activation technique, alongside PowerInfer, an open-source inference engine that dynamically distributes workloads across CPU and NPU. It supports one-click deployment of open-source models such as Qwen, DeepSeek, Llama, Phi, Mistral, and OpenAI GPT-OSS, as well as open source agents including ComfyUI, Flowise, OpenManus, and SillyTavern.
Founded in 2024, Tiiny AI Inc.’s team includes engineers from MIT, Stanford, Intel, Meta, HKUST, and SJTU, with feature rollout scheduled for CES January 2026.
