CryptoSlate
Tether has unveiled QVAC Psy, a family of foundational models framed as a "decentralized mind" that runs locally on phones, laptops, and peer-to-peer networks rather than centralized cloud servers. The stack ships with QVAC Fabric — a fork of llama.cpp spanning Linux, macOS, Windows, iOS, and Android — plus a unified SDK for fine-tuning, speech, OCR, and retrieval-augmented generation. Its first concrete model release, MedPsy, is a text-only medical LLM at 1.7B and 4B parameters, which Tether says scores 70.54 on seven closed-ended medical benchmarks against Google's MedGemma-27B-text-it at 69.95 — roughly seven times the size — and posts a 74.00 / 58.00 split on HealthBench and HealthBench Hard versus 65.00 / 42.67 for the Google model.
Why it matters
The move extends the mechanics of Tether's core business from money to intelligence. USDT already converts offshore dollar demand into a reserve stack dominated by short-duration Treasuries; in Q1 2026, Tether reported $1.04 billion in net profit and an $8.23 billion reserve buffer against roughly $183 billion in token-related liabilities. QVAC applies the same operating-cash flywheel to compute, models, and datasets — funding a long-duration infrastructure bet from the yield on the world's largest stablecoin rather than from venture rounds. The Asimov frame is more than branding: it positions AI as civilizational infrastructure, with the same permissionless premise that powered stablecoin adoption.
Market impact
The architectural bet is a different race from frontier labs. OpenAI, Anthropic, Google DeepMind, and xAI optimize for maximum general capability delivered through a centralized service; QVAC optimizes for deployability, privacy, latency, and survival when centralized services drop. Tether's SDK documentation says QVAC-powered apps continue working when the internet goes down, with peer-to-peer inference routed through the Holepunch stack. The credibility test is replication: MedPsy's strongest public benchmarks still come from Tether's own evaluation, the training corpus has not been released, and the model is text-only, English-only, and explicitly not for emergencies. If external researchers reproduce the gap, QVAC becomes the first credible example of a stablecoin issuer funding a competitive open-source AI stack; if the gap narrows, the infrastructure argument survives but the model claim weakens.
Frequently asked questions
-
What is Tether's QVAC Psy?
QVAC Psy is a family of foundational models Tether describes as 'rooted in the principles of Psychohistory,' paired with an edge-first runtime — QVAC Fabric, a fork of llama.cpp — that runs AI locally on consumer devices and across peer-to-peer networks.
-
How does MedPsy compare to larger medical models?
Tether reports MedPsy-4B scoring 70.54 across seven closed-ended medical benchmarks versus Google's MedGemma-27B-text-it at 69.95, and posting 74.00 / 58.00 on HealthBench and HealthBench Hard against 65.00 / 42.67 for the Google model — roughly seven times the parameter count.
-
Has MedPsy's performance been independently verified?
No. The strongest public benchmarks come from Tether's own evaluation harness; the training corpus has not been released, and the FAQ explicitly flags the model as text-only, English-only, unsuitable for emergencies, and vulnerable to hallucination.
-
How is Tether funding its AI push?
Q1 2026 operating results: $1.04 billion in net profit and an $8.23 billion reserve buffer against roughly $183 billion in token-related liabilities, with significant Treasury bill exposure. The same yield engine that funded the 8,888 BTC purchase in January is now backing QVAC.
-
What is the decentralization claim in QVAC?
QVAC decentralizes at the inference layer — users can run models locally, keep data on device, and route inference peer-to-peer via the Holepunch stack. Governance, model selection, and roadmap direction remain centrally coordinated by Tether.
WatcherGuru
Benjamin Cowen
Crypto News
CoinDesk
CoinTelegraph
WuBlockchain