Tether's 800M AI Agent Outperforms Models 2.5 Times Its Size in Six Tool-Use Tests
Key Takeaways
- •Tether introduced QVAC BotPsy on November 11, 2026, an open-weight 0.8 billion-parameter AI model built to complete real-world tasks by making the correct tool calls across multiple steps.
- •In benchmarks arranged by Tether, BotPsy outperformed every model with 1.2 billion or fewer parameters across six agentic tests: BFCL v4, ACEBench, τ-bench, Claw-Eval, BrowseComp-Plus, and ToolSandbox.
- •Tether claims BotPsy beats Alibaba's Qwen3.5-2B-Think, a model 2.5 times its size, scoring 46.1 versus 23.8 on BFCL multi-turn tool calling, and roughly doubled the results of its own base architecture, Qwen3.5-0.8B-Think.
- •Because it is based on Alibaba's open-source Qwen3.5 and small enough to run locally on personal hardware, BotPsy limits the sensitive information sent to central providers, an approach CEO Paolo Ardoino says gives users more control over how their personal data is processed.
- •BotPsy extends Tether's QVAC ecosystem, joining VisionPsy for visual question answering, TranscribePsy for speech, and TranslatePsy for translation as the company builds out its on-device AI lineup.
Tether has introduced QVAC BotPsy, an open-weight AI model built to execute real tasks through apps and external tools. Despite carrying only 0.8 billion parameters, the model outperformed larger systems on six agentic tests arranged by Tether — a result the company says points to the potential of purpose-built AI to serve everyday needs directly on individual devices. The announcement was published on November 11, 2026, alongside benchmark results comparing the model against larger competitors.
With the release, Tether — the company best known for issuing the USDT stablecoin — is pushing to bring more capable AI assistants onto personal devices and away from centralized servers, positioning small, specialized models as an alternative to cloud-based processing.
Tether shared the news through its official X (Twitter) account on November 11:
Tether's 800M AI Agent Beats Models 2.5x Its Size on Everyday Phone Tasks. Learn more:
— Tether (@tether) November 11, 2026
Outscoring Larger Models Across Six Benchmarks
According to Tether, QVAC BotPsy outperformed all models with 1.2 billion or fewer parameters on six agentic benchmarks: BFCL v4, ACEBench, τ-bench, Claw-Eval, BrowseComp-Plus, and ToolSandbox. Benchmarks of this kind score whether a model can chain tool calls across multiple steps to complete a task, rather than generate a text answer alone.
On four key metrics, BotPsy nearly doubled the scores recorded by its base architecture, Qwen3.5-0.8B-Think. Its BFCL v4 score climbed from 23.4 to 44.0, and its ACEBench result rose from 30.1 to 58.7. The model also posted 51.5 on τ-bench and 64.6 on ToolSandbox.
Tether also claims that BotPsy outperforms Alibaba's Qwen3.5-2B-Think, a model with 2.5 times the number of parameters. On BFCL multi-turn tool calling, BotPsy scored 46.1 points against the larger model's 23.8, while its base architecture managed 7.5 points.
In further comparisons cited by Tether, BotPsy significantly outpaced MiniCPM5-1B-Think with a score of 82.5, outshone Qwen3.5-2B-Think with 69.3, and edged out LLama3-3B-Think at 74.4.
The company framed the results as an indication that tailored AI can handle daily needs on individual gadgets rather than depending on larger, general-purpose systems. The figures come from comparisons arranged by Tether itself, though the model's open-weight release leaves room for outside developers to download it and run their own evaluations.
Turning Everyday Requests Into Actions
Unlike conventional AI models that focus on generating responses, BotPsy is built to make the right tool calls across multiple steps. Tether says this design could support practical tasks such as organizing activities, sending messages, and even arranging travel.
One ACEBench scenario illustrated the behavior. An initial attempt to send a message failed because of a lack of Wi-Fi connectivity and a full inbox. BotPsy detected both obstacles, activated Wi-Fi, cleared inbox space, and then delivered the message. A baseline model, by contrast, would typically retry the same failed action repeatedly.
The model can also determine when no available tool is needed to address a request, rather than making an unwanted call. Tether describes this capability as significant for assistants that must perform numerous actions and modify content.
Local AI Could Reduce Reliance on Centralized Servers
BotPsy is based on Alibaba's open-source Qwen3.5 and, at 0.8 billion parameters — the internal weights that make up a model — is small enough to run on personal hardware, meaning requests can be satisfied locally. That on-device processing limits the volume of sensitive information — emails, contacts, calendars, and more — that must be sent to central providers.
Tether CEO Paolo Ardoino noted that with local AI, users will have more control over the processing of their personal data and the accessibility of services for their assistant, underscoring the company's emphasis on on-device privacy.
BotPsy extends the QVAC ecosystem developed by Tether, a family of models that also includes VisionPsy, which tackles visual question answering; TranscribePsy, which addresses speech; and TranslatePsy, which handles translation tasks. Taken together, the lineup shows Tether building out its on-device AI push across visual, speech, and translation tasks.
Source: Crypto Ninjas
