Tether releases QVAC BotPsy for AI using tools on personal devices – Unite.AI

Tether AI Research announced on November 11, 2026 the release of QVAC BotPsy-0.8B, a 0.8 billion parameter open-weight model for tool calls on personal devices, and reported that it outperformed every model tested by up to 1.2 billion parameters across six agentic benchmarks.

Built on Alibaba’s Qwen3.5-0.8B, BotPsy specializes in function calls and multi-step agent workflows for local deployment. Its target tasks include calendars, messaging, files, device settings, web search, travel, commerce and customer operations, with low-to-medium difficulty jobs identified as the best fit. The surrounding application provides service connections, permissions, and execution controls.

In Tether’s ACEBench example, a message failed because Wi-Fi was turned off and the inbox was full. BotPsy enabled Wi-Fi and freed up space in the inbox before sending the message, while its base model repeated the failed action. The model can also recognize when none of the available tools satisfy a request, an important behavior for an assistant acting within real applications.

Specialization drives benchmark gains

In Tether’s model sheet evaluation, BFCL v4 went from 23.4 for the Qwen3.5-0.8B-Think base to 44.0 for BotPsy, ACEBench from 30.1 to 58.7, and ToolSandbox from 30.5 to 64.6. BotPsy also beat Qwen3.5-2B-Think on five of six agent benchmarks; the larger model brought up Claw-Eval, from 48.9 to 46.5. All five models were evaluated in thinking mode with a shared harness and the scores are percentages.

The biggest advantage concerns the continuity between the curves. Tether reports a BFCL v4 multi-spin score of 46.1 for BotPsy, versus 7.5 for its base and 23.8 for Qwen3,5-2B-Think. This test measures whether an agent can continue working when new information arrives, such as checking availability, handling a conflict, and updating a calendar. BotPsy’s 86.1 on the BFCL hallucination test measures a separate requirement: refrain from calling a tool when no available tool is suitable.

Tether’s broader results do not establish overall superiority. Its GPQA Diamond and MMLU-Pro scores of 38.1 and 48.9 for BotPsy rank well below the 59.9 and 65.5 of the 2B model. BotPsy’s BrowseComp-Plus score of 4.5 leads the pack of five models, but remains low in absolute terms, highlighting weak open-ended search performance. BotPsy is therefore designed to invoke tools for facts it does not know.

On-device training and deployment

A technical account describes four post-training phases: supervised fine-tuning, multi-teacher policy distillation, and two rounds of reinforcement learning. Three teachers specializing in Qwen3.5-4B covered the use of tools, general reasoning and following instructions during distillation. The reinforcement learning phases used testable, rule-based rewards, first emphasizing tool recall and then adding more math, coding, knowledge, and working with structured outcomes. The release omits training code and training data.

The resulting model also produces shorter reasoning traces than the baseline model. On GPQA Diamond questions that all five models answered correctly, Tether measured 3,690 reasoning tokens for BotPsy versus 11,563 for its base; on MMLU-Pro, the counts were 1,912 and 5,105. Because each model addressed these questions, the comparison isolates the length of the reasoning rather than the difficulty of the question, providing a useful indicator of efficiency for local inference.

The GGUF version provides three files for llama.cpp and QVAC SDK: BF16 at approximately 1.52 GB, Q8_0 at approximately 0.81 GB, and Q4_K_M at approximately 0.53 GB. Tether recommends Q8_0 after averaging 50.7 across ten benchmarks, close to the GGUF BF16’s 50.4 with about half the file size. Q4_K_M averages 48.9 and targets tighter memory budgets. The main repository provides BF16 safety tensor weights under the Apache 2.0 license.

Safety Limits and QVAC Plan

The limitations of the intended use of the model sheet are substantial. BotPsy was trained and tested primarily in English, which is the recommended language. Its Multi-IF evaluation covered seven other languages ​​and produced much weaker results in following instructions in Hindi and Russian. BotPsy makes mistakes and is not intended for security-critical or high-risk automated decisions, nor for applications involving money, personal information, health, and other sensitive topics.

On Tether’s automated red team across six damage categories, BotPsy produced unsafe responses 21% of the time, compared to 36% on its baseline. In the case of prompt-injection attacks, a particular concern for agents reading the output of untrusted tools, the reported unsafe rates were 31 percent and 46 percent, respectively. The smaller model improved in both tests while generating unsafe responses.

BotPsy joins QVAC’s vision, speech, health and translation models, with future development intended to combine their specialized capabilities. “Users should decide what their assistant can access and what external services it can connect to,” said Paolo Ardoino, CEO of Tether.

Tether’s long-term design uses an Android or iPhone interface for everyday interaction, outsources demanding tasks to a home PC or trusted peer-to-peer nodes, and routes web queries through encrypted peer-to-peer connections. That system is a stated vision; the current version provides the template and distribution files.

The documented strength of BotPsy is a specialized local action level for routine tasks. Whether it becomes a useful personal assistant depends on application-level permissions and execution controls, as well as planned Tether integration and hardware delegation that address the research, complexity, and security limitations documented in this release.