Mentioned in This Article
1. Tether AI unveils BotPsy, a small agent for personal devices
Tether, the company that issues the USDT stablecoin, on Wednesday, November 11, introduced QVAC BotPsy, a Tether AI agent with 800 million parameters built to carry out everyday tasks directly on phones and other personal devices. In the company's own testing, BotPsy posted the best result on five of six tool-use benchmarks, beating rival models with up to 2.5 times as many parameters.
The release adds a new piece to QVAC, the artificial intelligence program Tether has been building as an alternative to cloud-based assistants run by large technology companies. Unlike a chatbot that only produces text, BotPsy is designed to operate a device: changing settings, sending messages and managing schedules through a sequence of actions that it plans and carries out itself.
The central argument behind the launch is that a useful assistant does not necessarily require an enormous model. If a small system can reliably choose the right tools and recover from errors, it can run on hardware people already own, and the personal data it touches may never need to leave the device.
2. Why parameter count matters for on-device AI
Parameters are the internal numerical values a model adjusts during training, and together they determine what the model can represent and how capable it is. The largest frontier systems now contain parameter counts in the trillions, which is why they run in vast data centers rather than on a handset.
At 800 million parameters, BotPsy sits at the opposite end of that scale. A model of this size needs far less memory to load and run, which is the practical requirement for putting an assistant on a smartphone or laptop without a constant connection to remote servers. Fewer parameters generally translate into a smaller memory footprint, faster responses on modest chips and lower power draw.
The trade-off has traditionally been capability. Smaller models tend to make more mistakes, lose track of multi-step instructions and struggle to pick the correct function when several are available. Tether's results are an attempt to show that targeted training can close much of that gap for a specific job, in this case using tools on a device.
Running locally also changes the privacy equation. An assistant that manages messages, calendars and settings necessarily handles some of the most personal information on a device. When the model processes those requests on the phone itself, that information does not have to be transmitted to a remote server for interpretation, which is a central part of the case Tether is making for small models.
3. What makes BotPsy an agent rather than a chatbot
Tether describes BotPsy as agentic, meaning it is built to act rather than simply answer. Given a goal, the model works out which tools it needs, issues commands in the right order and changes course when a step does not succeed.
That last ability is often where small assistants fall down. A model that cannot recognize why an action failed tends to repeat it, producing a loop that never reaches the user's goal. An agent that can read the failure, identify its cause and fix the underlying problem behaves much more like a competent human operator.
In practice, this kind of agent is aimed at the routine, fiddly work people do on their phones every day: toggling connectivity, clearing space, writing and sending messages, and arranging calendar entries. Each of those tasks involves several distinct steps across different parts of the operating system, and each step can fail for reasons the user never sees.
4. How BotPsy handled a blocked message in testing
One benchmark scenario illustrates the difference. The model was asked to send a message on a device where Wi-Fi had been switched off and the inbox was already full. Completing the request required fixing two separate obstacles before the message could go out.
BotPsy turned Wi-Fi back on, freed up storage and then sent the message. The base model it was trained from, by contrast, kept attempting the same failing action without addressing either problem.
The task itself is modest, and nobody would call it a breakthrough on its own. But it captures the core quality that separates a dependable agent from an unreliable one: the capacity to diagnose an obstacle and remove it rather than repeating an error until the user gives up.
5. Built on Alibaba's Qwen3.5-0.8B
BotPsy is not trained from scratch. Tether started from Qwen3.5-0.8B, a compact open model from Alibaba's Qwen family, and gave it additional training focused specifically on tool use.
This approach, taking an existing general-purpose base model and specializing it, has become common across the AI industry because it lets developers concentrate their effort on a narrow skill. It also allows a clean comparison: by measuring BotPsy against the untouched Qwen3.5-0.8B, Tether can show how much its extra training changed the model's behavior.
The company also compared BotPsy with Qwen3.5-2B, a larger member of the same family. That model has roughly 2.5 times as many parameters, which is the basis for Tether's claim that its agent outperforms models more than twice its size.
6. BotPsy benchmark scores against larger Qwen models
On BFCL v4, a benchmark that tests whether a model can select and execute the right functions across a range of tasks, BotPsy scored 44.0%. The base Qwen3.5-0.8B managed 23.4%, and the larger Qwen3.5-2B reached 37.7%. The figures are weighted percentages calculated across several categories of tool use.
On ACEBench, which measures the accuracy of function calls, including whether the model picks the correct tool and fills in the correct arguments, BotPsy scored 58.7%. The base model scored 30.1%, while the 2-billion-parameter Qwen model scored 55.4%.
In total, Tether ran five models through six tool-use evaluations. BotPsy recorded the top score on five of them. Its average across all six came to 45.0%, compared with 38.8% for the larger Qwen3.5-2B.
The gains over the base model are the most striking part of the results. On both BFCL v4 and ACEBench, BotPsy roughly doubled the score of the model it was built from, suggesting that the additional training substantially changed how it handles tools rather than adding marginal polish.
7. Results come from Tether's own testing
All of the performance figures come from Tether's testing and its technical documentation. They have not been independently reproduced, and readers should treat them as the company's claims until outside researchers run the same evaluations.
That caveat applies to nearly every model launch in the industry, where developers routinely publish benchmark tables alongside new releases. Because BotPsy is distributed with open weights, however, independent testing is possible: anyone can download the model and run it against the same public benchmarks to check whether the numbers hold up.
Benchmarks also capture only part of the picture. Standardized tests measure performance on defined scenarios, while real devices present a far messier mix of apps, permissions and unexpected states. How closely BotPsy's benchmark advantage carries over to daily use is a question only broader deployment can answer.
8. What is QVAC?
QVAC is Tether's initiative to build artificial intelligence that runs on devices controlled by users instead of relying on centralized cloud infrastructure. The company first introduced it in May 2025, presenting it as an alternative to AI offered by large technology firms.
From the start, Tether said the program would eventually include agents capable of interacting with applications and potentially making payments in Bitcoin (BTC) and USDT. BotPsy, with its focus on operating apps and device settings, is a step toward the first half of that goal.
The program has grown steadily since then. In October 2025, Tether added a 41-billion-token dataset intended for training models in science and engineering, along with QVAC Workbench, an application for running models locally on a user's own machine. Earlier in 2026, the company released a software development kit for building AI applications that work offline on mobile and desktop platforms.
9. MedPsy and Tether's other compact models
BotPsy joins a growing set of small QVAC models aimed at specific jobs. The ecosystem already includes local models for medical use, translation and speech.
The medical model, QVAC MedPsy, is designed for consumer hardware. Its 1.7-billion-parameter version reportedly outperformed Google's MedGemma-27B on HealthBench Hard, a test that uses criteria set by physicians to grade responses in realistic clinical conversations. MedGemma-27B has roughly 16 times as many parameters, making that result another example of the pattern Tether is pursuing: narrowly trained small models competing with much larger general ones.
Taken together, the releases show a consistent strategy. Rather than trying to match the largest AI labs on raw scale, Tether is concentrating on compact, specialized models that can run privately on ordinary hardware.
10. Open-weight release and what comes next
BotPsy is available as an open-weight model under the Apache 2.0 license, which allows developers to download it, modify it and build it into their own products, including commercial ones. Tether is also offering compressed versions that reduce the amount of memory needed to run the model, widening the range of devices it can fit on.
There are limits to what has been released. The model card states that the training code and training data are not included, so outside developers can use and adapt the finished model but cannot fully reproduce how it was made.
Tether has not given a release date for a consumer assistant built on BotPsy. For now, the model is a building block for developers rather than a finished product, and its real-world value will depend on how well it performs once it is placed in front of everyday users and their unpredictable phones.