A startup called Tiiny AI showed up at CES 2026 with a device small enough to hold in one hand and a claim big enough to raise eyebrows: a 300-gram mini PC that runs 120-billion-parameter language models entirely on-device. At the time, Tiiny AI said it planned to launch the Pocket Lab on Kickstarter in February 2026 with a $1,399 super-early-bird tier.
The pitch is straightforward: buy the hardware instead of paying subscription or token fees, with model inference handled locally. In the same week, OpenAI began testing ads in ChatGPT for US users on the Free and Go tiers, while researchers described a prompt-injection-driven file-exfiltration chain in Anthropic’s Cowork research preview. Those developments make the appeal of local AI easy to understand.
But “appealing” and “proven” are different things. Here’s what we know, what we don’t, and what you should consider before backing this.
What’s Actually Inside
Tiiny AI lists the following specifications, which have not been independently tested:
- Processor: 12-core ARMv9.2 CPU with a custom NPU rated at about 190 TOPS
- Memory: 80GB LPDDR5X, which Tiiny AI estimated was worth more than $900 at then-current component prices
- Storage: 1TB SSD
- Power: 30W TDP, 65W typical system power
- Size: 14.2 x 8 x 2.53 cm, roughly 300 grams
- Record: Tiiny AI said Guinness had confirmed an upcoming certification for the smallest mini PC capable of running a 100B-parameter model locally
The company says the device runs a custom operating system called TiinyOS and supports models from the Qwen, DeepSeek, Llama, Phi, Mistral, and GPT-OSS families. It also claims support for frameworks including ComfyUI, Flowise, OpenManus, and SillyTavern.
The Software That Makes It Work
Running a 120B-parameter model on 80GB of memory within a 65W power envelope would require aggressive optimization. Tiiny AI says it uses two technologies based on published research.
TurboSparse is a neuron-level sparse activation technique. The paper reports that a modified 7B Mistral model left about 90% of its feed-forward neurons inactive, with 2.5 billion parameters active per token. Its modified 47B Mixtral model activated 4.3 billion parameters per token; mixture-of-experts routing and within-expert sparsity together left roughly 3% of the MoE-layer parameters active. Those results do not establish the active-parameter count or speed of Tiiny AI’s claimed 120B setup.
PowerInfer is the open-source inference engine behind the approach. The public implementation keeps frequently activated “hot” neurons on a GPU and computes less-used “cold” neurons on the CPU, reducing GPU-memory demand and CPU-GPU transfers. Tiiny AI says Pocket Lab adapts the approach to its CPU and NPU hardware, but the cited repository does not document that product-specific implementation.
The combination is plausible. Instead of computing every model parameter for every token, sparse inference exploits the fact that only part of the network is active at a time. The cited research does not independently validate Tiiny AI’s hardware.
What the Numbers Say (and Don’t Say)
At CES, Tiiny AI said it demonstrated the Pocket Lab running models at more than 20 tokens per second.
That could be practical for interactive chat, but Tiiny AI has not published head-to-head benchmarks against comparable setups.
The company claims performance “comparable to GPT-4o” and says the system covers “over 80% of real-world use cases.” These are marketing claims without published evaluation data to back them up. There are no standardized benchmark scores or comparisons against the same models running on other hardware. The reported Guinness certification concerns the size of a device capable of running a 100B model locally, not its output quality or speed.
The Privacy Angle
Local processing is the Pocket Lab’s main privacy pitch.
Tiiny AI says every query stays on the device and that there is no phone-home telemetry. If that holds in the shipping product, local execution would reduce the need to send conversation data to a cloud provider. It would not make the device physically incapable of leaking data through software, updates, networking, or a security flaw.
The supported list includes open-source and open-weight models, while the inference engine and sparse-activation research are published. That gives researchers more to inspect than a closed cloud service, but it does not verify Pocket Lab’s complete software stack.
How It Compares
The Pocket Lab doesn’t exist in a vacuum. Running AI locally is a well-established practice, and the alternatives are worth considering.
A Mac Studio with M4 Ultra (128GB unified memory, starting around $4,000) runs 70B models through Ollama or MLX at competitive speeds with better thermal headroom and a mature software ecosystem. It can’t fit in your pocket, but it also isn’t a Kickstarter project.
Ollama on existing hardware is free if you already have a machine with enough RAM. A used Mac Mini M2 Pro with 32GB can run capable 7B-13B models at interactive speeds for well under $1,000.
A high-end PC with an RTX 4090 (24GB VRAM) handles quantized 70B models and offers raw speed that the Pocket Lab likely can’t match, though at 4-5x the power draw and none of the portability.
The Pocket Lab’s 80GB of memory is its standout spec. That’s more than any consumer GPU and enough to fit quantized versions of genuinely large models. But 80GB of LPDDR5X running at ARM mobile speeds is a different beast than 80GB on an A100.
Red Flags to Watch
A few things give us pause.
The “OTA hardware upgrades” wording. TechRadar’s headline questioned Tiiny AI’s reference to over-the-air hardware upgrades. The accessible source text did not explain what the company meant, so the phrase should not be read as downloadable physical components.
Kickstarter delivery. Tiiny AI announced a $1,399 super-early-bird tier and said it planned to launch the campaign in February 2026. That is an offer for a crowdfunded product, not an independently tested retail device.
No independent benchmarks. The cited coverage does not include independent benchmark results. The available performance figures come from Tiiny AI’s demos and claims.
The 120B reality check. The cited materials do not provide enough detail about quantization format, runtime buffers, context length, or the exact model used to assess how a 120B model would fit in 80GB or perform over long sessions. The company’s 20-plus-tokens-per-second demonstration is a data point, not a comprehensive answer.
What This Means
Our analysis is that Pocket Lab points to a broader shift from DIY local-model setups toward dedicated consumer hardware. The techniques Tiiny AI cites - sparse activation and heterogeneous inference - are documented in published papers and open-source code.
But the gap between “technically possible” and “practically useful” is where most hardware startups live and die. At $1,399 on Kickstarter, this is a bet on a team’s ability to deliver hardware that matches their software ambitions.
If you’re already running local models and want dedicated hardware, the Pocket Lab is worth watching. If you’ve never run a local model, an existing Mac or Linux box with Ollama will teach you more for less money and less risk.
The Bottom Line
The Tiiny AI Pocket Lab has published specifications, ties to open-source research, and a potentially useful local-processing design. What it does not have yet is independent validation of its performance or privacy behavior. Until someone outside the company runs their own tests, the 120B-in-your-pocket story remains an impressive claim, not a proven product.