The outfit says its Pocket Lab can run models that usually demand hardware which costs more than a second-hand car.

The startup says its Pocket Lab measures just 14.2 by 8 by 2.53 centimetres and weighs about 300 grams, although on paper it claims it can deploy a 120-billion-parameter model and handle what it calls PhD-level reasoning. It is aimed at punters who want local LLMs without having to flog a kidney for a data centre rig.

The firm’s spec sheet lists an ARM v9.2 twelve-core CPU, a custom heterogeneous module delivering about 190 TOPS, 80GB of LPDDR5X and a 1TB SSD. Tiiny AI insists the tiny slab can run a 120-billion-parameter model entirely on-device, thanks to a mix of aggressive quantisation and what it describes as two clever tricks.

The first is TurboSparse, which is a neuron-level sparse activation system that boosts inference efficiency without gutting the model’s intelligence. The second is PowerInfer, an open-source heterogeneous inference engine with more than 8,000 GitHub stars. Tiiny AI said it dynamically distributes workloads across the CPU and NPU, aiming to deliver server-grade performance at a fraction of the energy draw.

The Pocket Lab can run models from GPT OSS, Llama, Qwen, DeepSeek, Mistral, and Phi, and the firm claims full offline operation with one-click deployment across dozens of open-source LLMs and agent frameworks. The bundle carries a 30W TDP, though Tiiny AI says the typical load is closer to 65W.

The outfit intends to roll out the gadget at CES 2026, though it has not said when it will ship or how much it will cost. Tiiny AI is punting it as the world’s smallest supercomputer, although the supercomputing crowd might need a stiff drink before they accept that a pocket-sized ARM slab now counts as one.