Field Guide FG-02 / Pre-arrival profile / Rev 2026.10.06
A computer around power-bank size that promises serious offline inference. The idea is unusually prosumer-friendly. The evidence is not ours yet. So this page separates what Tiiny says, what follows from the architecture, and what the parcel still has to prove.
Every attractive sentence gets an evidence label
| Spec / claim | Evidence | Prosumer translation | |
|---|---|---|---|
| 14.2 × 8 × 2.53 cm · ~300 g | Vendor | → | If the shipping unit matches the claim, this is genuinely jacket-scale compute — the jacket test, finally. The full kit still includes power and cables. |
| ARMv9.2 · custom SoC | Vendor | → | Efficiency first. Also: no CUDA comfort blanket. The software ecosystem matters as much as the silicon. |
| dNPU · ~190 TOPS | Vendor | → | Potentially excellent watts-per-token — but only for models the vendor toolchain can convert. You are buying the converter pipeline too. |
| 80 GB LPDDR5X · 1 TB NVMe | Vendor | → | Huge capacity for the size. The important unanswered question is bandwidth and how the CPU/NPU memory split behaves in real workflows. |
| ~30 W AI workload / ~65 W system ceiling | Vendor | → | Exactly the kind of power envelope an always-on private companion wants. Wall-meter verification comes after delivery. |
| “Runs up to 120B” | Vendor | → | Fits is only question one. Question two: which 120B, what quantisation, what context, and how many tokens per second? |
The doctrine is compelling even before the benchmark
Chat, note search, voice and quick questions are all-day, low-intensity jobs. A small efficient node can own the interactive day while a bigger GPU machine sleeps.
The pitch is private, local inference with encrypted storage and no token meter. That is exactly the right promise for this category — and exactly what we should verify packet by packet.
One device present all day, another waking for heavy work. Different boxes doing the jobs that fit their shape instead of one giant machine doing everything badly.
Pre-arrival classification; bars are not measurements
Fine print before the pledge button
If a great new open model arrives and the dNPU converter does not, excellent silicon can still be a generation behind. We need to measure converter cadence as part of product maturity.
Capacity, bandwidth and CPU/NPU partitioning are different facts. The brochure gives the cup; we still need to inspect the straw.
Model, quant, context, throughput, first-token latency, power and sustained behaviour. One headline number cannot answer all seven.
APIs, local files and weights should remain useful if the company, desktop app or conversion service changes. Long-term autonomy is a feature, not an apocalypse fantasy.
No pass badges before our units arrive
| Test | Pre-arrival call | Evidence | What we will actually do |
|---|---|---|---|
| First result | Wait | Vendor | Time unboxing → first useful answer on our own PDF. Count every install, login and hidden step. |
| Reading | Wait | Vendor | Run exact 30B/70B/120B models; capture first-token latency, sustained tok/s and context. |
| Library | Wait | Unverified | Measure dB(A), surface temperature and sustained speed after twenty minutes. |
| Backpack | Likely | Inferred | Weigh the complete useful kit, not just the 300 g chassis; count charger and cables. |
| Offline | Wait | Vendor | Disconnect WAN and inspect outbound packets, activation checks and local workflow completeness. |
| Meeting room | Wait | Inferred | Run a private document workflow from another screen with the internet unavailable. |
| Electricity | Wait | Vendor | Wall-meter idle, interactive and sustained loads; cost the 24/7 case. |
| Next door | Wait | Unverified | Check API portability, model-converter cadence, Olares/network interoperability and vendor independence. |
300 grams of private AI is a strong sentence. It is not yet our result.
That is why this page is a pre-arrival Field Guide rather than a review. When the units arrive, every attractive claim above either graduates to OBSERVED / MEASURED, gets qualified, or fails. All three outcomes are useful.