Tether Says BotPsy Beats Larger Qwen AI With 800M Parameters

Tether, the firm behind USDT, introduced Tether BotPsy on November 11, 2026. This 800-million-parameter artificial intelligence agent is engineered to operate on personal hardware and execute multi-step tasks offline, without requiring a cloud connection. Developers can acquire the model weights immediately and integrate them into their own applications.

According to Tether, BotPsy outperformed a larger Qwen competitor during its internal tool-use evaluations. The debut is part of the company’s QVAC initiative, which is developing a suite of on-device AI tools controlled by users rather than centralized platforms.

BotPsy Targets Tasks That Require Several Steps

While most small language models only answer queries, BotPsy was trained to take action—selecting appropriate software tools, running commands sequentially, and recovering when a process fails mid-task. That recovery capability is distinct: in Tether’s test scenario featuring disabled Wi-Fi and a full inbox, the base model got stuck retrying the broken step in a loop, whereas BotPsy turned on the Wi-Fi, cleared storage, and sent the message.

A finalized consumer-facing AI phone assistant is not yet available. The weights are accessible for developers, though Tether has not announced a release date for a polished end-user application.

Developers Get BotPsy Weights and Compressed Versions

The model is distributed under the Apache 2.0 license, permitting developers to download, alter, and deploy it freely. Compressed versions are provided to lower memory demands for local deployment. As noted in the model card, Tether withheld both the training data and the training code.

Earlier in 2026, Tether launched an SDK designed for building offline AI applications across mobile and desktop systems, and BotPsy builds upon that framework.

BotPsy Beats Larger Qwen Model in Tether’s Tests

As reported by Decrypt, BotPsy achieved a score of 44.0% on BFCL v4, a benchmark evaluating software-function selection and execution. Its underlying base model, Qwen3.5-0.8B, achieved only 23.4%. Meanwhile, the larger Qwen3.5-2B model—possessing more than double the parameter count—scored 37.7%, meaning BotPsy surpassed it by over six percentage points.

On ACEBench, which measures tool selection and function-call arguments, BotPsy recorded 58.7%, compared to 30.1% for its base model and 55.4% for the 2-billion-parameter Qwen. Across five models and six tool-use assessments, BotPsy ranked first in five categories with an average score of 45.0%, outperforming the larger Qwen’s average of 38.8%.

QVAC’s Broader Plans Include Bitcoin and USDT Payments

Tether launched QVAC in May 2025 with the objective of creating agents that interact with apps and eventually facilitate Bitcoin and USDT payments. Payments remain on the roadmap rather than functioning as an active BotPsy feature. In October 2025, the project expanded with a 41-billion-token dataset focused on science and engineering training, alongside the release of QVAC Workbench to allow local model execution.

Additionally, Tether created QVAC MedPsy for medical AI tasks on consumer hardware. Its 1.7-billion-parameter model outperformed Google’s MedGemma-27B on HealthBench Hard, a benchmark testing realistic clinical interactions designed around physician-formulated criteria.

Tether Faces a Separate Senate Records Deadline

The rollout arrives amid heightened regulatory scrutiny. US Senator Richard Blumenthal has requested that Tether and three other cryptocurrency companies surrender records regarding USDT utilization by sanctioned entities, as well as communications with Nxum, a consulting firm co-founded by former White House crypto adviser Bo Hines. The firms must begin turning over documents by November 30, with all materials due by January 4, 2027.

Blumenthal cautioned that “alternative measures to obtain compliance” could follow if the January deadline is missed. Tether has previously pushed back against similar pressure; when a September Senate staff report claimed USDT played a part in Iran’s shadow banking operations, the company stated it had assisted authorities in freezing nearly $550 million in Iran-associated tokens throughout 2026.

FAQ

Which Alibaba model is BotPsy built on?

BotPsy uses Alibaba’s Qwen3.5-0.8B architecture as its foundation.

What does BotPsy’s Apache 2.0 license allow developers to do?

It allows developers to download, modify, and ship the model freely in their applications.

What did BotPsy do when a messaging task encountered disabled Wi-Fi and a full inbox?

It enabled Wi-Fi, cleared enough storage, and successfully completed the message.

What are the compressed BotPsy versions intended to reduce?

They are meant to cut memory requirements for local deployment.

Which earlier QVAC tool lets users run AI models locally?

QVAC Workbench, introduced in October 2025, lets users run models locally.

This article is for information only and is not financial advice.