A 4-Billion-Parameter Medical AI Just Beat Google's 27-Billion on a Phone

A 4-billion-parameter medical AI just beat Google's 27-billion-parameter MedGemma on the standard U.S. medical-licensing exam set, and the smaller model runs on a phone. Tether, the company behind the USDT stablecoin used by roughly 573 million wallets, built it. The benchmark numbers are Tether's own; the deployment math is where the comparison starts to bite the cloud-AI incumbents.

Tether's QVAC MedPsy-4B scores 70.54 on a seven-benchmark closed-ended medical evaluation, edging Google's MedGemma-27B-text-it at 69.95, according to the company's release on Hugging Face. The gap widens where the harder tests live: HealthBench Hard (58.00 vs. 42.00, +16.00), the full HealthBench (74.00 vs. 65.00), and MedXpertQA (30.61 vs. 25.18). The smaller sibling, MedPsy-1.7B, averages 62.62, beating Google's MedGemma-1.5-4B-it (51.20) by more than 11 points at less than half the parameters, and matching Qwen3-4B-Thinking-2507 at roughly 2.4x the size advantage in the other direction. Every number in this paragraph is Tether-reported, run on Tether Data's own evaluation harness; independent reproduction has not yet appeared in the public packet.

MedPsy-4B produces roughly 909 output tokens where the comparable Qwen3 model needs 2,953, a 3.2x reduction, because the model is trained on a narrow medical domain instead of the whole web. The quantized Q4_K_M footprint lands at about 2.6 GB; the 1.7B version is about 1.2 GB, which is what an $80 Android phone in Lagos, Karachi, or Manila can hold in RAM. Eight benchmark suites back the score: MedQA-USMLE, MedMCQA, MMLU Health, MMLU-Pro Health, MedXpertQA, PubMedQA, AfriMedQA, and HealthBench. The weights ship under Apache 2.0 through the QVAC SDK. A clinician can run a USMLE-grade assistant on the device in her hand, with no data center in the loop and no patient note leaving the building.

That is also the unit-economics case Ardoino made on Eye on AI this week. His argument: a vendor selling a $200 annual AI subscription is delivering between $1,000 and $5,000 of inference cost per customer, and the gap is being absorbed privately while the company is venture-funded. The moment one of those vendors goes public, or has to defend its margin to a public-market investor, the bill either lands on the subscriber or on retail shareholders. Ardoino frames it as a parallel to the early USDT playbook: give the user the product for free, monetize elsewhere. The 573 million wallets already on Tether's rails are the customer base the on-device thesis is being run against.

The comparison exposes a unit-economics story. On-device inference for a narrow model is cheaper per call than the cloud for a large generalist, because the marginal cost of running a 4-billion-parameter model on a phone that is already paid for is electricity, while the marginal cost of running a 27-billion-parameter model in a data center is the next H100 cluster, the next substation, and the next ten-year depreciation schedule. The trillion-dollar GPU build-out of 2024 to 2026 was priced against a world in which every useful AI call goes through a hyperscaler. If the on-device curve crosses first, that capex is on a depreciation clock the cloud is not.

The counterforce is also named. Hyperscalers can drop inference pricing below the on-device break-even if they have to. Closed medical data and regulated clinical workflows are not solved by a phone-resident model alone. Every benchmark in Tether's release is self-reported, not independently reproduced. The reader should treat the +11.42 and the +16.00 as a vendor's claim, not as a settled fact, until a hospital IT team or an independent lab reruns the eval on the published weights. Ardoino's five-year, 90%-on-device forecast is a CEO bet, not a measured adoption rate. The phone in your pocket is now running the numbers, though, and the GPU order book is the one that should be paying attention.