Tether's tiny BotPsy agent beats bigger models at running tasks on your phone

What's going on

Tether is best known as the issuer of USDT, the world's largest stablecoin (a cryptocurrency pegged to the US dollar). For a while now, though, the company has also been building its own AI division under the name QVAC. On Wednesday, November 11, 2026, its research arm, Tether AI Research, released QVAC BotPsy-0.8B. It's a small AI model with 800 million parameters, built to handle everyday tasks on personal devices like smartphones.

BotPsy isn't meant to be a classic chatbot that just answers questions. It's an agent, which means it's supposed to take action on its own. It works out which tools an app offers, runs commands in the right order and changes course when something goes wrong. The weights are out as an open-weight model under the Apache 2.0 license. Developers can download the model, modify it and build it into their own apps, including commercial ones. There are also compressed versions that need less memory.

A small model with specialized training

BotPsy is based on Qwen3.5-0.8B, a small language model from Alibaba. Tether fine-tuned this base model specifically to work with software tools. The industry terms for this are tool use or function calling. The model doesn't just write text. It also generates function calls like "turn on Wi-Fi" or "move appointment to 3 p.m.", and the app then actually carries them out.

For comparison, Decrypt notes that today's largest frontier models have trillions of parameters. BotPsy is tiny by comparison, and that's the point. A model this small runs directly on your phone or laptop, so personal data like messages, calendar entries and settings never has to go to a cloud server.

Tether gives an example from the ACEBench benchmark. An assistant has to send a message, but Wi-Fi is off and the inbox is full. BotPsy turned on Wi-Fi, freed up storage and sent the message. The base model just kept retrying the same failed action. Decrypt calls this a "fairly mundane feat", but says these are exactly the kinds of problems agents have to solve once they leave the chat window.

Tether says BotPsy can also tell when it needs a tool at all, when a task takes several steps and when none of the available tools fits. The company lists rescheduling appointments, setting reminders and troubleshooting phone problems as use cases.

The numbers: benchmarks according to Tether

Tether tested five models on six agent benchmarks: BFCL v4, ACEBench, τ-bench, Claw-Eval, BrowseComp-Plus and ToolSandbox. Among other things, these tests check whether a model calls the right function with the right arguments, including across multiple conversation turns. BotPsy scored best in five of the six tests. It averaged 45.0 percent, compared with 38.8 percent for the larger Qwen3.5-2B. According to the Spanish site Crypto Economy, BotPsy beat every tested model with 1.2 billion parameters or fewer. It also edged out Qwen3.5-2B-Think, the variant with a reasoning mode.

On the four main tests, BotPsy roughly doubled the base model's score. The biggest jump came on multi-turn tasks (BFCL Multi-Turn), where the base model managed just 7.5 points. On the telecom section of τ-bench, which simulates support conversations, BotPsy scored 82.5 points. On BFCL's hallucination test it scored 86.1 points, the highest of any model Tether tested. That test measures whether a model makes up functions or calls one when nothing suitable exists.

One important caveat: all of these numbers come from Tether's own tests. Crypto Economy points out that they don't prove BotPsy is better across AI tasks in general. It was only compared with small models and doesn't go up against large cloud models.

Part of the QVAC ecosystem

Tether introduced QVAC in May 2025 as an alternative to the AI services of big tech companies. The idea is that AI should run on devices users control themselves instead of in centralized data centers. According to Decrypt, Tether's long-term plan is autonomous agents that operate apps and might even pay with Bitcoin and USDT.

In October 2025, Tether added a 41-billion-token dataset for science and engineering models, plus QVAC Workbench, an app for running AI models locally. In early 2026 came an SDK, a developer toolkit for offline AI apps on mobile and desktop. BotPsy belongs to the "Psy" family, which also includes VisionPsy (image understanding), TranscribePsy (speech recognition), TranslatePsy (translation) and MedPsy. Tether says the 1.7-billion-parameter version of MedPsy beat Google's much larger MedGemma-27B on HealthBench Hard. That test scores realistic doctor conversations against criteria developed by physicians.

There's no consumer-ready assistant built on BotPsy yet, and Tether hasn't given a date. According to the Model Card, training code and training data aren't part of the release.

What you need to try it

With 800 million parameters, BotPsy is very light. At full precision (BF16) it takes up about 1.6 GB. With 4-bit quantization, which stores the model's numbers at lower precision, it's roughly under 1 GB, or about 2 GB with some headroom. You don't need a graphics card: the model runs smoothly on regular CPUs, Apple Silicon Macs and Google Colab. 8 GB of RAM and a few gigabytes of free disk space are plenty. In practice, though, BotPsy only shows what it can do when you give it tools as function definitions, for example through an agent framework. As a plain chatbot, it's too small for demanding conversations.

The bigger picture

BotPsy fits a clear trend: instead of ever-larger all-rounders, small models are being specialized for individual tasks. If an 800-million-parameter model calls tools more reliably than a 2-billion-parameter one, that matters for smartphones, where memory, battery life and privacy all count. The Apache 2.0 license makes it easy for developers to adopt.

There are limits, though. All the benchmarks come from Tether, and the only comparisons are with small Qwen models. It's an open question how BotPsy stacks up against other specialized tool-use models, and how it holds up in real apps with messy data. A 44 percent score on BFCL v4 also means a lot of tasks still fail. That's not yet a solid foundation for an assistant that changes settings on your phone by itself. And without training data and training code, nobody can reproduce the results. Still, it's notable that a crypto company is now regularly shipping competitive small models.