Answers from the documentation; checking which model is on duty
Hi - I answer from the OpenSmartRoute documentation: routing, the API, plans and quotas, self-hosting. Ask away, or open a support ticket if you need a person.
Grounded in the docs - follow a source before acting on it.
Tether releases QVAC BotPsy, a tool-using AI model for personal device - OpenSmartRoute
Introduction - what was announced and why it matters
Tether AI Research announced the release of QVAC BotPsy-0.8B on November 11, 2026. It is a new AI model designed to call tools on personal devices. This model is smaller than many others but performs better on specific tasks. It matters because it shows how smaller models can be more effective for certain uses. This can help reduce costs and improve local deployment of AI assistants.
The release highlights progress in making AI models more capable at handling multi-step workflows. It also demonstrates that a smaller model can outperform larger ones on important benchmarks. This can influence how AI is built and used in everyday devices. For engineers, it offers a new tool to improve local AI functions. Managers can see it as a step toward more efficient and capable AI assistants.
Model details - size, architecture, and training process
QVAC BotPsy-0.8B has 0.8 billion parameters. Parameters are the parts of a model that learn from data. The model is built on Alibaba’s Qwen3.5-0.8B architecture. This means it uses a specific design that helps it understand and generate language. The training process involved several steps. First, supervised fine-tuning was used to teach the model basic skills.
Next, the model went through multi-teacher on-policy distillation. During this stage, three specialist teachers helped it learn to use tools, reason, and follow instructions. After that, two rounds of reinforcement learning improved its ability to call tools correctly. These reinforcement stages used rules and rewards that could be checked for accuracy. The training data and code for this process are not publicly available.
The model was trained mainly in English. It was tested in other languages, but with weaker results in some. The training focused on making the model good at calling functions and managing multi-step workflows. The design emphasizes local deployment, meaning it can run on personal devices without needing cloud access. This makes it suitable for tasks that require privacy and quick responses.
Benchmark performance - scores and comparison with other models
Mercor announced a $750 million funding round, valuing the company at $20 billion. The money will boost infrastructure and research for AI model evaluation and training.
QVAC BotPsy-0.8B outperformed all tested models up to 1.2 billion parameters on six agentic benchmarks. These benchmarks measure how well a model can handle tasks like reasoning, tool use, and multi-turn conversations. In Tether’s ACEBench test, BotPsy scored much higher than its base model. Its ACEBench score was 58.7, compared to 30.1 for the base.
It also did better on other benchmarks. For example, its scores on BFCL v4, ToolSandbox, and other tests showed significant improvements. On BFCL v4, its score rose from 23.4 to 44.0. On ToolSandbox, it went from 30.5 to 64.6. These scores are percentages, indicating how often the model answered correctly. BotPsy beat larger models on five of six benchmarks, except on Claw-Eval, where a 2-billion-parameter model led.
A key strength of BotPsy is its ability to maintain context across multiple turns. Its multi-turn score was 46.1, much higher than the 7.5 of its base model. This means it can keep track of ongoing tasks better. It also scored 86.1 on a hallucination test, showing it can refrain from calling tools when no suitable ones are available. This ability is important for real-world applications, where accuracy and safety matter.
However, the model’s scores on broader knowledge tests like GPQA Diamond and MMLU-Pro were lower than larger models. Its GPQA score was 38.1, and MMLU-Pro was 48.9, compared to larger models that scored higher. Its open-ended research performance remains weak, with a low BrowseComp-Plus score of 4.5. This indicates it is mainly designed for calling tools and managing workflows, not for open-ended research.
The training process involved multiple stages. After supervised fine-tuning, the model was distilled using three specialized teachers. Reinforcement learning then improved its tool use, reasoning, and instruction-following skills. Rewards were rule-based and verifiable, focusing first on tool calling, then on math, code, and structured output. The training process aimed to make the model efficient and accurate for local deployment.
Capabilities and limitations - what BotPsy can and cannot do
BotPsy is designed to call tools for routine tasks on personal devices. It can handle workflows involving calendars, messaging, files, device settings, web research, travel, commerce, and customer operations. Its strength lies in low- to medium-difficulty work that requires multi-step reasoning. For example, it can enable Wi-Fi, free inbox space, and send messages automatically.
The model recognizes when no available tool fits a request. This is an important feature for acting within real applications. It can also keep context across multiple interactions, making it suitable for ongoing tasks. Its ability to refrain from unnecessary tool calls helps prevent errors and unsafe responses.
However, BotPsy has limitations. Its performance on broad knowledge tests is weaker than larger models. It is not optimized for open-ended research or high-stakes decisions. Its open-ended research score remains low, and it is not intended for safety-critical applications. The model was mainly trained in English, with weaker results in other languages like Hindi and Russian. It can make mistakes, especially in complex or unfamiliar situations.
Safety is a key concern. In testing, BotPsy produced unsafe responses 21% of the time, which is better than its base model. Under prompt-injection attacks, unsafe responses occurred 31% of the time, compared to 46% for the base. While smaller, it still generates unsafe outputs, highlighting the need for safeguards. Its design emphasizes calling tools for facts it does not know, rather than generating uncertain answers.
The model is best suited for local deployment where permissions and controls can be managed. It relies on application-level safeguards to prevent misuse. Its low reasoning token count indicates it produces concise explanations, which can improve efficiency. Still, it is not a general-purpose research model but a specialized assistant for routine, multi-step workflows.
Application examples - real-world tasks and workflows
BotPsy can be used in various everyday tasks that involve multiple steps. For example, it can manage calendars by checking availability, updating appointments, and sending reminders. It can handle messaging workflows, such as enabling Wi-Fi, clearing inboxes, and sending messages automatically.
In web research, BotPsy can look up information, verify facts, and compile summaries. It can assist with travel planning by checking schedules, booking tickets, and updating itineraries. For commerce, it can track orders, update shopping lists, and manage payments. Customer service workflows include responding to inquiries, updating records, and routing requests.
These capabilities make BotPsy useful for personal assistants on devices like smartphones or tablets. It can also be integrated into larger systems for business workflows. Its ability to call tools and manage multi-step processes helps automate routine tasks. This reduces manual effort and speeds up responses.
The model’s design allows it to recognize when a task cannot be completed with available tools. This prevents it from giving false or unsafe answers. It can also adapt to different workflows by connecting to various services and permissions. This flexibility makes it suitable for a range of personal and professional applications.
Developers can build custom workflows around BotPsy’s tool-calling abilities. They can connect it to local apps, web services, or other tools. Its ability to handle multi-turn interactions makes it suitable for complex tasks. This can improve productivity and user experience in many settings.
Safety and ethical considerations - risks and safeguards
Safety remains a critical concern with AI models like BotPsy. During testing, it produced unsafe responses in about one-fifth of cases. This shows that even smaller models can generate problematic outputs. The model’s unsafe response rate under prompt-injection attacks was also notable. It occurred in nearly one-third of such cases, indicating vulnerabilities.
To address these risks, BotPsy was designed to recognize when no suitable tool exists for a request. It refrains from calling tools unnecessarily, reducing the chance of unsafe actions. Its ability to identify when it cannot fulfill a request helps prevent errors. However, it still makes mistakes, especially in complex or unfamiliar situations.
The model was mainly trained in English, with weaker instruction-following results in other languages. This highlights the importance of language-specific testing and safeguards. Its low open-ended research performance means it should not be used for high-stakes decisions involving health, money, or personal data. Users should implement application-level permissions and controls to limit unsafe outputs.
OpenSmartRoute can help manage safety by routing requests through models with different safety profiles. It can enforce rules that prevent sensitive data from leaving on untrusted models. It also measures routing accuracy and flags requests that might pose risks. This layered approach helps ensure safer deployment of models like BotPsy.
Developers and users should continuously monitor outputs and update safeguards. Regular testing and feedback loops can reduce unsafe responses. Combining technical safeguards with user oversight helps mitigate risks. AI models for routine tasks should be used with caution, especially in sensitive contexts.
How OpenSmartRoute helps
OpenSmartRoute is an open-source router that manages AI requests. It sends each request to the best-fit model, agent, or tool from a predefined catalog. The team can set how requests are scored based on quality, cost, speed, and safety. This helps balance performance and safety for each request.
Requests are pinned with rules to keep data safe. For example, personal data stays on an on-premises model, and region or cost limits are enforced. The router learns from outcomes, giving more traffic to models that perform well. It also reduces traffic to models that fail or produce unsafe results. This helps improve overall system safety and efficiency.
Adding a new model is simple. It only requires updating the catalog with the new entry. The system then starts routing requests to it, competing with existing models. This makes it easy to test and deploy new AI capabilities without changing the app. The router also tracks costs and performance, helping teams optimize their AI use.
By measuring routing accuracy with tools like 'osr eval', teams can ensure requests are handled correctly. The platform provides a clear view of costs and safety metrics. It also includes input guards to prevent prompt injections and protect personal data. This setup helps teams deploy models like BotPsy safely and efficiently.
What to do - next steps for users and developers
Users interested in BotPsy should explore how it can improve routine workflows on personal devices. They can test its ability to call tools and manage multi-step tasks. Developers can integrate BotPsy into existing systems by connecting it to local apps and services. They should also consider safety measures, such as permissions and safeguards.
For teams using OpenSmartRoute, updating the catalog with new models like BotPsy is straightforward. They can set routing rules to prioritize safety and performance. Regularly testing routing accuracy and safety metrics helps maintain system quality. Using tools like 'osr eval' can identify issues early.
Both users and developers should stay informed about new model releases and improvements. They can participate in testing and provide feedback to enhance safety and effectiveness. Continuous monitoring and updates ensure AI assistants remain reliable and safe for everyday use.
Implementing layered safeguards and managing permissions are key steps. They help prevent unsafe outputs and protect sensitive data. As models evolve, keeping the routing system flexible and transparent supports better AI deployment.
Pierre Fabre and XtalPi announced a partnership to discover new small-molecule drugs. XtalPi provides AI and robotics, while Pierre Fabre handles selection and development.
