Tether Releases QVAC BotPsy-0.8B Open-Weight AI Model for Local Devices

Tether AI Research launched QVAC BotPsy on Nov. 11, 2026, releasing an open-weight 0.8-billion-parameter artificial intelligence model designed to execute multi-step tool calls on consumer hardware ¹. Built atop Alibaba's Qwen3.5-0.8B architecture, the compact model targets local execution to perform routine mobile assistant tasks, including scheduling, messaging, and system management, without transferring personal data to cloud environments ¹². The launch coincides with Tether's broader push into decentralized machine learning infrastructure, which previously saw Tether Capital Backs AIsa to Expand Agentic AI Payments Infrastructure.

Evaluation benchmarks published alongside the open-source weights show BotPsy-0.8B roughly doubled the agentic capacity of its base model across core industry tests ¹. On the BFCL v4 evaluation, the model scored 44.0 compared with 23.4 for Qwen3.5-0.8B-Think, while reaching 58.7 on ACEBench and 64.6 on ToolSandbox ¹. In five out of six evaluated benchmarks—including BFCL v4, ACEBench, τ-bench, BrowseComp-Plus, and ToolSandbox—the 0.8B model exceeded the accuracy of the 2.5-times-larger Qwen3.5-2B-Think, trailing only on Claw-Eval with a score of 46.5 against 48.9 ¹². Furthermore, on BFCL v4 multi-turn tasks, BotPsy achieved a score of 46.1 compared with 23.8 for the 2B model, while scoring 86.1 on tool hallucination prevention tests ¹².

Local Architecture and Distributed Strategy

According to Tether, BotPsy addresses computational limits on edge devices by specializing in function calling rather than broad factual retrieval, where large parameter counts typically retain advantages ¹². In multi-turn customer troubleshooting tests via τ-bench telecom, the model scored 82.5, surpassing both Qwen3.5-2B-Think (69.3) and MiniCPM5-1B-Think (74.0) ¹². Tether CEO Paolo Ardoino stated that local deployment reduces the dependency on centralized providers when handling sensitive data such as calendars and emails, granting users direct authority over device permissions ¹. The model weights are published under an Apache 2.0 license in BF16 format, with quantized GGUF variants for deployment in llama.cpp and QVAC tooling ².

Tether plans to position BotPsy as an operational agent alongside its other domain models, including VisionPsy, TranscribePsy, MedPsy, and TranslatePsy ¹. The project's long-term framework envisions mobile devices delegating complex tasks, such as document analysis and travel planning, over encrypted peer-to-peer networks to home personal computers or decentralized nodes ¹. At the time of writing, $USDT traded at $0.9989 amid a broader cryptocurrency market capitalization of $2.46 trillion.