World's smallest AI supercomputer fits in your pocket

What if you could carry an AI supercomputer that runs a 120-billion-parameter model completely offline inside your jacket pocket?

That is exactly what US-based deep-tech startup Tiiny AI just introduced. The company unveiled the Tiiny AI Pocket Lab on December 10, and Guinness World Records has already verified it as the world’s smallest personal AI supercomputer.

Power bank size, data center ambitions

The device looks almost identical to an everyday portable power bank. Yet, it operates entirely without internet access, external cloud servers, or desktop-class graphics cards.

Under the hood, the Pocket Lab features a 12-core ARMv9.2 CPU running at just 65 watts of power. Despite that tiny thermal footprint, it can host models between 10B and 100B parameters, which Tiiny AI says handles over 80 percent of daily tasks. It can even stretch to run massive 120B models locally for multi-step reasoning.

To pull off this kind of efficiency on lightweight silicon, the system relies on two main software technologies. First is TurboSparse, which activates only essential neurons during inference without degrading model smarts. The second is Powerinfer, a popular open-source project with over 8,000 GitHub stars that spreads AI workloads intelligently across the CPU and NPU.

One-click open-source ecosystem

Hardware is only half the battle, so Tiiny AI built an entire local ecosystem around it. The device supports one-click installs for leading open weights, including Llama, Qwen, DeepSeek, Mistral, Phi, and GPT-OSS.

It also streamlines agent workflows out of the box. Users can deploy OpenManus, ComfyUI, Flowise, and SillyTavern without manual setup headaches. All files and context history stay behind bank-level local encryption, giving users private long-term memory. The company plans to showcase the hardware at CES in January 2026.

Why it matters

Cloud AI dominance has pushed developers and sensitive enterprise users into a tricky privacy corner. For researchers, legal teams, and creators handling strictly confidential material, Pocket Lab represents a rare self-contained sanctuary that fits in a backpack.

We think Tiiny AI made a very clever bet by leveraging sparse activation and hybrid CPU-NPU scheduling instead of waiting for cheaper desktop GPUs. It tackles the energy crisis of big data centers at a micro level.

However, the elephant in the room is token generation speed. Squeezing a 120B model into a 65W enclosure is an impressive engineering feat, but if the output crawls at single-digit tokens per second, practical utility will be narrow. Unless you have an absolute requirement for air-gapped reasoning right now, we think it makes sense to see hands-on benchmarks at CES before pulling the trigger.