Tether Just Built an AI That Knows When to Say ‘I Can’t Do That’

CN
3 hours ago

Key Takeaways

  • Tether released QVAC Botpsy on Nov. 11, an 800-million-parameter AI assistant that runs locally on devices.
  • Botpsy beat Alibaba’s 2.5x larger model on four benchmarks and is now publicly available on Hugging Face.
  • Tether plans to connect its QVAC models so phones can offload heavy tasks to home computers via encrypted links.

Stablecoin issuer Tether has released a new artificial intelligence model designed to run directly on a phone or personal computer. The company says the model, called QVAC Botpsy, is small but effective at completing tasks, and in its own tests, it outperformed rival models more than twice its size.

According to the Nov. 11 announcement, Tether built the model to act as a personal assistant that handles tasks such as booking trips, moving meetings, and setting reminders, rather than merely answering questions.

The company tested Botpsy, which has 800 million parameters, on six well-known benchmarks that measure an AI’s ability to use apps and tools. The company says it outperformed every other tested model with 1.2 billion parameters or fewer. On four of the primary tests, it also scored higher than Qwen3.5-2B-Think, a model from Chinese tech giant Alibaba that is 2.5 times larger.

One of the most significant gains came in tests requiring multi-step, back-and-forth actions, such as rescheduling a meeting after a calendar conflict arises. Botpsy scored 46.1 on this evaluation, nearly double the 23.8 scored by the larger Alibaba model. The base model it was built from managed just 7.5.

The model also excelled at recognizing its own limits. When none of the available tools matched a request, it was better at declining to act rather than hallucinating an answer. It scored 86.1 on a test measuring this capability, the highest among the models Tether compared.

“An assistant that manages your calendar, messages, and documents needs access to some of your most personal information,” said Paolo Ardoino, Tether’s CEO. “Running AI locally can reduce the need to send that information to centralized providers, giving people greater control over how it is processed.”

He argued that users should control what their assistant can see and which external services it can access, and that developers should be free to run and modify these models without relying on a single company.

Botpsy is the latest in a family of AI models from Tether’s QVAC research project. Others in the family handle images, speech, health inquiries, and translation. Tether says it will explore how these models can work together. The company’s long-term vision is for a smartphone to act as an always-on companion.

For heavy computational workloads, such as analyzing a lengthy document or planning a complex trip, the phone would offload the work to a home computer and receive the results back. Web searches could likewise be routed through private, encrypted connections between users’ own devices instead of corporate servers.

Botpsy is “open-weight,” meaning its files are publicly available. Researchers and developers can download, study, and build upon it. It is available now on Hugging Face, a popular repository for AI models.

免责声明:本文章仅代表作者个人观点,不代表本平台的立场和观点。本文章仅供信息分享,不构成对任何人的任何投资建议。用户与作者之间的任何争议,与本平台无关。如网页中刊载的文章或图片涉及侵权,请提供相关的权利证明和身份证明发送邮件到[email protected],本平台相关工作人员将会进行核查。

Share To
APP

X

Telegram

Facebook

Reddit

CopyLink