Tether Releases 800M-Parameter BotPsy Agent, Beats Larger Qwen

Tether released BotPsy on November 11, 2026, an 800-million-parameter AI agent built to run on personal devices and execute multi-step tasks without a cloud connection. The company published the model weights for developers immediately, while withholding training code and training data, per the model card.

Tether's own benchmarks claim BotPsy outperforms a larger competitor. The model scored 44.0% on BFCL v4, a benchmark testing software-function selection and execution. Its base architecture, Alibaba's Qwen3.5-0.8B, managed 23.4%. The Qwen3.5-2B — a model with more than twice the parameters — reached 37.7%, meaning BotPsy cleared it by more than six points.

On ACEBench, which evaluates tool selection and function-call arguments, BotPsy scored 58.7%, against 30.1% for its base model and 55.4% for the 2-billion-parameter Qwen. Across five models and six tool-use evaluations, BotPsy led on five and averaged 45.0%; the larger Qwen averaged 38.8%.

What separates BotPsy from a standard chatbot?

Most small language models answer questions. Tether trained BotPsy to act: pick the right software tool, run commands in sequence, and recover when a step fails mid-task.

The difference shows in the company's test scenario involving disabled Wi-Fi and a full inbox. The base model kept retrying the failed step in a loop. BotPsy enabled Wi-Fi, cleared storage, and finished sending the message. That recovery behavior, rather than raw parameter count, drove the benchmark gap.

What do developers actually get?

The release ships under the Apache 2.0 license, so developers can download, modify and integrate the model into their own applications without licensing friction. Compressed versions accompany the weights to cut memory requirements for local deployment.

Consumer users will wait. Tether has not set a launch date for a polished end-user product, positioning the current release as a developer toolkit. Earlier in 2026, the company released an SDK for building offline AI apps across mobile and desktop operating systems; BotPsy adds an agent layer to that stack.

Where does BotPsy sit in Tether's QVAC strategy?

BotPsy sits inside QVAC, the initiative Tether introduced in May 2025 to build on-device AI that users, not platforms, control. The roadmap includes agents that interact with applications and eventually make Bitcoin and USDT payments — payments remain a plan, not a shipped feature.

The initiative expanded in October 2025 with a 41-billion-token dataset aimed at science and engineering training, plus QVAC Workbench, an application that lets users run models locally. Tether also built QVAC MedPsy for medical AI on consumer hardware; its 1.7-billion-parameter version outperformed Google's MedGemma-27B on HealthBench Hard, a benchmark of realistic clinical conversations built around physician-developed criteria.

Why does the launch carry regulatory baggage?

The release lands while Tether faces a separate Senate records deadline. US Senator Richard Blumenthal has demanded that Tether and three other crypto firms hand over records on USDT use by sanctioned entities and on the company's dealings with Nxum, a consulting firm co-founded by former White House crypto adviser Bo Hines.

Firms must begin producing documents by November 30, with everything due by January 4, 2027. Blumenthal warned of "alternative measures to obtain compliance" if the January deadline slips. Tether has pushed back before: after a September Senate staff report alleged USDT's role in Iran's shadow banking network, the company said it had helped authorities freeze nearly $550 million in Iran-linked tokens in 2026.

For now, the open-weight release extends QVAC's developer footprint without committing Tether to a consumer launch date, while the document-production window that opened November 30 gives Blumenthal's committee its next enforcement checkpoint.