BotPsy: On-Device Personal Assistant with a Sub-Billion-Parameter Agentic

Tether AI Research launched BotPsy on November 11, 2026, releasing an open-weight, 0.8-billion-parameter language model tuned for on-device agentic workflows. Built on Qwen3.5-0.8B, the system acts as a personal assistant for daily tasks across email, calendar management, web research, finance, and system operations. Details in Hugging Face Blog's report show the model roughly doubles the tool-use accuracy of its base architecture on consumer hardware.

This report examines BotPsy's benchmark performance, multi-stage training pipeline, efficiency optimizations, and deployment formats for engineers and privacy-focused practitioners. The architecture addresses key bottlenecks in local agent deployment, including reasoning loops and third-party cloud dependencies. Readers evaluating small agent frameworks will find a breakdown of how the model matches or exceeds systems more than twice its size.

01Benchmark gains across six agentic evaluations

On standard benchmarks, the model roughly doubles the capability of its base model, Qwen3.5-0.8B-Think. Scores on the Berkeley Function Calling Leaderboard version four rose from 23.4 to 44.0, while ACEBench increased from 30.1 to 58.7. On tau-bench across retail, telecom, and banking subsets, the model posted a mean score of 51.5 compared to 23.2 previously.

Against rival models in the sub-1.2-billion-parameter category, BotPsy recorded top scores across all six evaluation suites. The model outperformed competing systems including LFM2.5-1.2B-Think and MiniCPM5-1B-Think on ToolSandbox, Claw-Eval, and BrowseComp-Plus. These results establish the architecture as a leading lightweight engine for constrained edge hardware.

Evaluations against Qwen3.5-2B-Think, an architecture 2.5 times larger, demonstrated consistent advantages across five of six test suites. The 0.8-billion model beat the 2-billion alternative on BFCL version four by 44.0 to 37.7 and on ACEBench by 58.7 to 55.4. It also achieved higher marks on ToolSandbox with 64.6 over 41.2, trailing only on Claw-Eval.

02Multi-turn workflows and tool-calling decisions

In complex multi-turn evaluations, the model maintained policy adherence across extended conversational interactions. On tau-squared-bench telecom interactions, the architecture reached 82.5 compared to 69.3 for Qwen3.5-2B-Think. Multi-turn tasks within BFCL version four rose from 7.5 on the base model to 46.1, nearly double the 23.8 mark recorded by the 2-billion-parameter model.

Domain breakdowns showed leading multi-step accuracy across commercial and administrative sectors. The agent achieved a score of 58.3 in commerce and services tasks, surpassing the 47.4 of the next closest model. In device and system management workflows, it achieved 46.1 compared to 31.9 for its closest competitor.

Knowing when to avoid tool calls proved critical during negative testing. On the BFCL version four hallucination benchmark, the system scored 86.1 by declining to invoke tools when matching functions were absent. On ACEBench special cases involving incomplete requests or invalid parameters, accuracy rose from 13.7 on the base model to 51.3.

03Reasoning efficiency and prompt-injection safety

Automated red-teaming across six harm categories demonstrated safety gains over baseline and competing small architectures. The system generated unsafe responses in 21 percent of test interactions, improving from 36 percent in its base model. That performance beat LFM2.5-1.2B-Think at 35 percent and MiniCPM5-1B-Think at 47 percent, while prompt-injection vulnerabilities fell to 31 percent.

Execution efficiency improved through concise internal deliberation that reduced runtime latency. On correctly answered GPQA Diamond questions, the system solved problems using 3,690 reasoning tokens, compared to 11,563 tokens required by the baseline model. Similarly, MMLU-Pro tasks required 1,912 tokens rather than 5,105 tokens, delivering a 2.7-fold reduction in computational overhead.

Repetitive deliberation errors showed dramatic decreases under demanding benchmark evaluations. Instances where the model became trapped in repetitive reasoning loops during IFBench tests fell from 22.9 percent to 1.4 percent. Furthermore, test runs reaching the hard 32,000-token context limit dropped from 28.3 percent down to 0.2 percent during edge inference.

04Four-stage training recipe and specialist distillation

The development team created the assistant through a structured four-stage post-training pipeline initialized from Qwen3.5-0.8B weights. The sequential process advanced through supervised fine-tuning, multi-teacher on-policy distillation, and two successive rounds of reinforcement learning. Each phase relied on the highest-performing checkpoint from the preceding stage to reinforce tool-calling capabilities.

During the distillation stage, the student network received guidance from three distinct specialist Qwen3.5-4B teacher models. Each teacher model had undergone targeted reinforcement learning focused on a specific operational skill before transferring token probability distributions. This knowledge transfer paired imitation of curated trajectories with rule-based tool-calling rewards to reinforce execution accuracy.

This training regimen simultaneously reinforced underlying reasoning and broad subject comprehension. Instruction-following marks on IFEval jumped from 59.7 to 90.8, while IFBench advanced from 22.0 to 55.9. General knowledge benchmarks showed parallel progress, with GPQA Diamond rising from 33.2 to 38.1 and MMLU-Pro advancing from 41.6 to 48.9, outpacing all tested sub-1.2-billion models.

05Local ecosystem integration and peer-to-peer delegation

Tether Data designed the software architecture for open distribution under the Apache 2.0 license for research and educational purposes. Release assets include Hugging Face safetensors alongside quantized GGUF binaries in bf16, Q8, and Q4 formats. These runtime targets ensure compatibility with llama.cpp and the QVAC SDK across smartphones and local workstations.

Long-term plans for the project center on decentralized infrastructure and local fine-tuning environments. The company plans to connect the agent to Tether Hypersearch, enabling decentralized peer-to-peer web queries that bypass corporate data trackers. Additionally, the QVAC Fabric mechanism will support local customization, allowing the assistant to adapt to specific user workflows without cloud synchronization.

System expansions also include distributed compute delegation during network state changes. Future iterations will allow mobile devices to offload demanding tasks to local desktop machines or trusted peer nodes during offline-online transitions. Secure agent-to-agent protocols will further enable collaborative execution across local area networks while ensuring user credentials and personal files remain strictly on owned hardware.