Tether's 800M-Parameter BotPsy Beats Models 2.5x Its Size

Tether, the company behind the stablecoin USDT, on Wednesday unveiled QVAC BotPsy, an 800-million-parameter AI agent built to carry out everyday tasks on personal devices, according to Decrypt. In the company's own testing, the relatively small model outperformed competitors with up to 2.5 times as many parameters, a result Tether says suggests AI assistants may not need enormous models to get things done. The release marks the latest expansion of Tether's QVAC initiative, which the stablecoin issuer first introduced in May 2025 as an alternative to AI services run by large technology companies, and it arrives as Tether continues building out an AI ecosystem designed to run on user-controlled devices rather than centralized cloud infrastructure, per Decrypt's reporting.

Event: What BotPsy Is and How It Performed

Unlike conventional chatbots that simply answer questions, BotPsy is described by Tether as agentic, meaning it is designed to take action rather than just respond. According to Decrypt, the model can identify which tools an application needs, execute commands in the correct order, and adjust its approach when something goes wrong. Tether illustrated this with a benchmark scenario in which an assistant needed to send a message but encountered two obstacles: Wi-Fi was disabled and the inbox was full. Per Decrypt's account of Tether's testing, BotPsy turned on Wi-Fi, cleared enough storage, and completed the task, while the underlying base model repeatedly attempted the same failed operation without adapting.

BotPsy is built on Alibaba's Qwen3.5-0.8B architecture, which Tether further trained to specialize in interacting with software tools, according to Decrypt. Its 800 million parameters sit far below the trillions of parameters that frontier AI models typically carry today, and a smaller parameter count generally translates into lower memory requirements for running the model locally. On BFCL v4, a benchmark that evaluates an AI system's ability to select and execute software functions across different tasks and conversations, BotPsy scored 44.0%, according to Tether's technical documentation as reported by Decrypt. By comparison, the underlying Qwen3.5-0.8B model scored 23.4%, while the considerably larger Qwen3.5-2B reached 37.7% on the same test. On ACEBench, which evaluates the accuracy of function calls including the selection of appropriate tools and arguments, BotPsy scored 58.7%, versus 30.1% for its own base model and 55.4% for the larger Qwen3.5-2B, per Decrypt. Tether tested five models across six tool-use evaluations and found BotPsy posted the highest results on five of them, with an average score of 45.0% across those tests compared with 38.8% for the larger Qwen model, Decrypt reported.

Implications for Tether's AI Strategy

BotPsy is the latest addition to QVAC, and its release continues a pattern Tether has followed throughout 2025 and into 2026 of pairing stablecoin dominance with a parallel push into decentralized, on-device artificial intelligence. Decrypt notes that Tether expanded the QVAC initiative in October 2025 with a 41-billion-token dataset intended for training science and engineering models, alongside QVAC Workbench, an application that lets users run AI models locally rather than through cloud servers. Earlier in the year, per Decrypt, Tether released a software development kit giving developers tools to build offline AI applications across both mobile and desktop operating systems. The company also developed QVAC MedPsy, a medical AI system designed to run on consumer hardware; its 1.7-billion-parameter version reportedly outperformed Google's considerably larger MedGemma-27B on HealthBench Hard, a test of realistic clinical conversations evaluated using physician-developed criteria, according to Decrypt's reporting.

Tether's longer-term ambitions for QVAC, as described in Decrypt's coverage, extend beyond task automation to autonomous agents capable of interacting with applications and potentially making payments using Bitcoin and USDT. BotPsy is currently available as an open-weight model under the Apache 2.0 license, meaning developers can download, modify, and integrate it into their own applications, Decrypt reported. Tether has also released compressed versions intended to further reduce memory requirements for local deployment, though the company has not announced a release date for a finished consumer assistant powered by BotPsy, and its published model card indicates that the training code and training data are not included in the release.

CoinHours Analysis: Reading the Benchmark Gap

CoinHours analysis: the benchmark spread Decrypt reported is notable less for the raw scores than for what it implies about efficiency in agentic AI design. A model with 800 million parameters outperforming one with 2 billion parameters on both BFCL v4 (44.0% versus 37.7%) and ACEBench (58.7% versus 55.4%) suggests that targeted fine-tuning for tool-use tasks can matter more than sheer scale, at least within the narrow domain of executing device commands. This matters for Tether specifically because QVAC's stated goal is running capable AI on personal hardware without relying on centralized cloud infrastructure; a smaller model that performs competitively is directly aligned with that lower-memory, on-device objective. No CoinHours-saved market dataset was applicable to this story, since BotPsy's release is a product and infrastructure event rather than a price-moving market development, and this analysis is confined to interpreting the figures Decrypt attributed to Tether's own technical documentation rather than asserting any causal link to USDT's market position.

What to Watch

Several open questions remain following the release. Tether has not specified a timeline for a finished consumer assistant built on BotPsy, leaving the gap between an open-weight research release and a shipped product undefined, per Decrypt. The company's model card also states that training code and training data were withheld from the release, which limits independent verification of the benchmark claims by outside researchers. Whether BotPsy's agentic capabilities extend into Tether's stated ambition of autonomous Bitcoin and USDT payments, rather than remaining limited to device housekeeping tasks like toggling Wi-Fi or clearing storage, will be a key signal of how QVAC's broader roadmap is progressing. Developers now have access to the Apache 2.0-licensed weights and compressed variants, so community testing and independent benchmarking in the weeks ahead may offer a check on Tether's own reported figures.

Further Reading

For background on related crypto infrastructure concepts, see CoinHours' guide on crypto wallets explained and the comparison of hot wallet vs cold wallet storage, relevant to the device-based, user-controlled model QVAC is pursuing.

Further reading / context