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 across the AI hardware world: a 300-gram mini PC that runs 120-billion-parameter large language models entirely on-device, no cloud required. The Pocket Lab is now headed to Kickstarter this month with a $1,399 price tag.
The pitch is straightforward. Pay once, own your AI forever. No subscriptions, no token fees, no data leaving your device. In a week where OpenAI started putting ads in ChatGPT and Anthropic shipped a product with known security vulnerabilities, the appeal of completely local AI has never been clearer.
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
The hardware specs are genuinely impressive for a device this size:
- Processor: 12-core ARMv9.2 CPU with a custom NPU delivering ~190 TOPS
- Memory: 80GB LPDDR5X (standalone market value exceeds $900, according to Tiiny AI)
- Storage: 1TB SSD
- Power: 30W TDP, 65W typical system power
- Size: 14.2 x 8 x 2.53 cm, roughly 300 grams
- Record: Guinness World Records verified it as “The Smallest MiniPC (100B LLM Locally)”
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 supports agent 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 requires aggressive optimization. Tiiny AI leans on two key technologies, both rooted in academic research from Shanghai Jiao Tong University.
TurboSparse is a neuron-level sparse activation technique. The core insight: during inference, most neurons in a large model’s feed-forward layers don’t activate for any given input. TurboSparse uses a custom activation function called dReLU to push sparsity to extreme levels - 90% in dense models and up to 97% in mixture-of-experts architectures. Instead of running the full 120 billion parameters, the device only computes the 2.5 to 4.3 billion that actually fire for each token.
PowerInfer is the open-source inference engine (8,000+ GitHub stars) that orchestrates the actual computation. It profiles which neurons activate frequently (“hot” neurons) and keeps those on the NPU for fast access. Less-used “cold” neurons get computed on the CPU. This heterogeneous approach avoids the bottleneck of shuttling data between processors.
The combination is clever. Instead of brute-forcing a massive model through limited hardware, these tools exploit the statistical structure of how language models actually work during inference.
What the Numbers Say (and Don’t Say)
At CES, Tiiny AI demonstrated the Pocket Lab running models with “real-world decoding speeds of 20+ tokens per second.” The company’s broader claim is 18-40 tokens per second depending on the model.
For context, 20 tokens per second is roughly the speed of a fast human typist. It’s usable for interactive chat. It is not, however, anywhere close to what a cloud GPU cluster delivers, and Tiiny AI hasn’t 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. No standardized benchmark scores. No comparisons against the same models running on other hardware. No independent verification beyond the Guinness record, which only confirms the device can run a 100B model locally - not how well it runs.
The Privacy Angle
This is where the Pocket Lab’s value proposition gets interesting regardless of raw performance numbers.
Every query stays on your device. There’s no telemetry phone-home (according to Tiiny AI’s claims). No conversation data training anyone else’s model. No ads being targeted based on what you ask your AI assistant. At a time when ChatGPT now serves ads personalized by your conversation topics and where major AI tools are shipping with prompt injection vulnerabilities, a device that physically cannot leak your data has real appeal.
The supported model list - Qwen, DeepSeek, Llama, Mistral, Phi - is entirely open-source or open-weight. You can inspect what’s running. The inference engine is open-source. The sparse activation research is published. This is about as transparent as an AI product gets.
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” claim. Tiiny AI has mentioned over-the-air hardware upgrades, which is physically impossible. You cannot download more RAM. This likely refers to firmware and software optimizations that improve how the hardware is utilized, but the misleading terminology suggests marketing that’s ahead of engineering precision.
Kickstarter delivery. The estimated delivery date is August 2026. Hardware Kickstarters have a well-documented history of delays, specification changes, and outright failures. A pre-order deposit of $1,299 locks in the lowest price, but it also means paying for a product that doesn’t exist yet from a company founded in 2024.
No independent reviews. As of this writing, no tech publication has done a hands-on review with their own benchmarks. Everything we know about performance comes from Tiiny AI’s own demos and claims.
The 120B reality check. Running a 120B model in INT4 quantization on 80GB of memory is mathematically feasible. Running it well - with acceptable context length, coherent long outputs, and reasonable latency - is a different question. The 20 tokens/second CES demo is a data point, not a comprehensive answer.
What This Means
The Pocket Lab represents a real trend: local AI hardware is becoming a product category, not just a DIY project. The underlying technology - sparse activation, heterogeneous inference engines - is legitimate research with published papers and working 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 real specs, real open-source technology, and a real privacy advantage. What it doesn’t have yet is real independent benchmarks. Until someone outside the company runs their own tests, the 120B-in-your-pocket story remains an impressive demo, not a proven product.