Tether Unveils 800-Million-Parameter AI Agent BotPsy, Built for Everyday Tasks on Phones and Other Devices, Beating a 2.5x Larger Model in Tether's Own Tests

What happened

According to Decrypt, Tether has released QVAC BotPsy, an AI agent with 800 million parameters designed to run everyday tasks on personal devices. Decrypt describes Tether as the company behind USDT, the world's largest stablecoin. Decrypt explains that parameters are the values a model learns during training; fewer parameters generally mean less memory is needed to run it. Decrypt also notes that "agentic" AI does not just answer questions like a traditional chatbot but takes action: it works out which tools are needed, runs commands in the right order, and adjusts its approach when something goes wrong.

According to the Hugging Face model card, BotPsy-0.8B is built for function calling and agentic workflows. Function calling means letting the AI operate features inside apps, such as adding an event to a calendar in the model card's example. It is built on Qwen3.5-0.8B, is suited to deployment directly on devices, and is released under the Apache 2.0 license. The model card describes it as a "personal assistant" that uses tools to complete simple tasks such as web search, calendar, email and messaging, and device control.

Decrypt reports that BotPsy can handle multi-step tasks involving phone settings, messaging and scheduling, potentially without sending personal information to cloud servers. Decrypt relays a test scenario described by Tether: an assistant has to send a message, but Wi-Fi is off and the inbox is full. BotPsy turns on Wi-Fi, frees enough space and completes the task, while its underlying base model keeps retrying the same failing action.

Benchmark results: where the small model wins

Benchmarks are standardized tests used to compare AI model capabilities. For example, Decrypt explains that BFCL v4 evaluates how well an AI selects and runs software functions across different tasks and conversations, while ACEBench evaluates function-calling accuracy, including choosing the right tool and parameter values. Decrypt says Tether tested five models on six tool-use evaluations, and BotPsy scored highest on five of them, with the results coming from the company's own testing. According to the model card, all models were tested in "thinking mode" (reasoning before answering) using the same evaluation framework. All figures below come from the model card, as percentage scores; higher is better.

According to the model card, Qwen3.5-2B-Think leads on Claw-Eval (48.9 vs. 46.5). On BrowseComp-Plus, BotPsy is the highest of the five models, but the model card admits the score is low in absolute terms. It trails clearly on knowledge benchmarks: 38.1 vs. 59.9 on GPQA Diamond and 48.9 vs. 65.5 on MMLU-Pro. The model card says knowledge still largely scales with parameter count, and that BotPsy is designed to call tools for facts it does not know rather than serve as a knowledge model.

Limitations and what to know before use

• According to the model card, BotPsy works best on low-to-medium-difficulty tasks and still makes mistakes.
• According to the model card, it is not intended for safety-critical or high-stakes automated decisions, nor for applications involving sensitive topics such as money, personal information or health.
• According to the model card, it was trained and tested mainly in English, and English is recommended; in Multi-IF tests across seven other languages it mostly performs well, but Hindi and Russian scores are much lower.
• According to the model card, no tested model scored above 5 on τ³-banking, the banking subset of the τ-bench tests.
• According to the model card, the training code and training data are not part of this release.

What it means for general readers

Decrypt argues this suggests AI assistants may not need huge models to get work done: if small models can reliably handle routine jobs such as settings, messaging and scheduling, such assistants could run on personal devices. However, Decrypt notes that Tether has not announced a release date for a full consumer assistant powered by BotPsy, so ordinary users cannot yet use a finished product.

For developers, Decrypt notes that BotPsy is offered as an open-weight model under the Apache 2.0 license, meaning the model itself can be downloaded, modified and integrated into their own applications. According to the model card, GGUF files (BF16, Q8_0 and Q4_K_M versions) are also available for local-run tools such as llama.cpp and the QVAC SDK. Decrypt says Tether has also released compressed versions to lower memory needs for local deployment.

Decrypt notes that Tether first launched QVAC in May 2025, positioning it as an alternative to AI services run by big tech companies. In October 2025, Tether released a 41-billion-token dataset and QVAC Workbench; earlier in 2026 it released an SDK. The company has also developed QVAC MedPsy, a medical AI system. BotPsy is the newest addition along this path.

Browse the latest news in this topic

• Crypto & blockchain