TL;DR
- Tether AI Research released BotPsy-0.8B, an 800-million-parameter open-weight model designed for local AI agents that can execute multi-step phone and tool tasks.
- In Tether’s testing, BotPsy beat Qwen3.5-2B-Think despite being 2.5 times smaller and roughly doubled its base model’s scores across four core agentic benchmarks.
- The model also led tested systems on BFCL hallucination checks and joins Tether’s broader QVAC ecosystem for specialized on-device AI applications and personal assistants locally.
Tether AI Research has released QVAC BotPsy-0.8B, an open-weight artificial intelligence model designed to turn everyday requests into multi-step tool calls on personal devices. According to the official announcement, the model contains 800 million parameters and was tested across six agentic benchmarks. Tether says BotPsy outperformed every tested model with 1.2 billion parameters or fewer and beat Qwen3.5-2B-Think despite being 2.5 times smaller. The release expands the company’s work around AI agents and local execution.
BotPsy Targets Multi-Step Tool Use on Personal Devices
BotPsy was evaluated on BFCL v4, ACEBench, τ-bench, Claw-Eval, BrowseComp-Plus and ToolSandbox. On four core evaluations, it roughly doubled the scores of its underlying Qwen3.5-0.8B-Think architecture. BFCL v4 rose from 23.4 to 44.0, ACEBench from 30.1 to 58.7, τ-bench from 23.2 to 51.5 and ToolSandbox from 30.5 to 64.6. The comparison suggests specialization can materially improve tool use without increasing model size. The approach fits Tether’s broader QVAC local AI strategy, which emphasizes user-controlled hardware.
The model is intended for tasks that require several actions rather than a single response. Tether highlighted examples such as rescheduling meetings, creating reminders and troubleshooting phone problems. In one ACEBench test, BotPsy turned Wi-Fi back on and cleared inbox space before sending a failed message, while its base model repeatedly retried the same action. BotPsy is designed to recognize when tools should be used, when multiple steps are required and when no available tool is appropriate. Its compact architecture supports the same on-device direction as Tether’s QVAC SDK.
Tether reported additional gains in multi-turn and telecom tasks. BotPsy scored 46.1 on BFCL multi-turn tool calling versus 23.8 for Qwen3.5-2B-Think and 7.5 for its base model, while its τ-bench telecom score reached 82.5. It also achieved 86.1 on BFCL hallucination tests, the highest among models Tether evaluated. The published results are Tether’s benchmark findings and do not establish universal superiority across every AI workload. BotPsy joins VisionPsy, TranscribePsy, MedPsy and TranslatePsy in the QVAC ecosystem, extending Tether’s on-device AI model lineup. Its model weights and evaluation details are available, allowing researchers to adapt the system. Tether’s longer-term vision is for BotPsy to power a personal QVAC companion, with smartphones handling everyday interactions and home computers or trusted peer-to-peer nodes processing more demanding requests. The company says encrypted P2P connections could also route web searches without relying on corporate cloud infrastructure.
