A 0.8B on-device assistant: BotPsy beats a model 2.5 times its size on five of six tool-use benchmarks
Synopsis
Tether AI Research released BotPsy, a 0.8B-parameter model built on Qwen3.5-0.8B for function calling and agentic workflows that runs on-device. Trained through supervised fine-tuning, multi-teacher on-policy distillation and two reinforcement-learning rounds, it lifts BFCL v4 from its base's 23.4 to 44.0 and τ-bench from 23.2 to 51.5, and beats Qwen3.5-2B-Think, a model 2.5 times its size, on five of six tool-use and agentic benchmarks, trailing only on Claw-Eval (46.5 vs 48.9). It still makes mistakes, targets low-to-medium difficulty tasks and is not intended for high-stakes use.
Interpretation
A model under one billion parameters is trained specifically as a tool-calling personal assistant, handling search, calendar, email and device control locally instead of sending requests to a cloud model. The prior route relies on frontier cloud LLMs with high and fluctuating operating costs, and third-party providers can modify, deny or deprecate access at any time, while many routine tasks do not need frontier-level intelligence. Across six tool-use and agentic benchmarks, BotPsy-0.8B averages 45.0, against 24.7 for its base Qwen3.5-0.8B-Think, 38.8 for the 2.5-times-larger Qwen3.5-2B-Think, 25.3 for LFM2.5-1.2B-Think and 22.7 for MiniCPM5-1B-Think.
Training runs in four sequential stages: learning from demonstrations, learning from larger specialist teachers, then two reinforcement-learning rounds driven by rule-based verifiable rewards. Supervised fine-tuning raises tool use but pushes general knowledge and instruction following below the base model, and the later stages have to recover them. Supervised fine-tuning, multi-teacher on-policy distillation and two RL rounds run in sequence, each starting from the previous stage's best checkpoint; the distillation stage routes each prompt to one of three Qwen3.5-4B teachers, each RL-trained for one skill, using per-token reverse KL divergence as the signal.
Tool-use gains come without a cost to general knowledge or instruction following, which end up above the base model. The base Qwen3.5-0.8B-Think scores 59.7 on IFEval, 22.0 on IFBench, 33.2 on GPQA Diamond and 41.6 on MMLU-Pro. BotPsy scores 90.8, 55.9, 38.1 and 48.9 on those four, above every other tested model of 1.2B parameters or fewer, though Qwen3.5-2B-Think remains clearly higher on knowledge (59.9 and 65.5).
Perspective
Evaluation covers six tool-use and agentic benchmarks, two knowledge benchmarks, three instruction-following benchmarks, multi-turn dialogue and app-grouped scenarios; all models were evaluated in thinking mode, and benchmarks using external APIs were run on November 3, 2026. The model is primarily trained and tested in English and is positioned for research and educational use on low-to-medium difficulty tasks.
Open-ended retrieval remains weak: BrowseComp-Plus at 4.5 is the best of the five models but low in absolute terms, and complex information-retrieval scenarios score 15. BFCL web search (14.0) and memory (16.3) trail Qwen3.5-2B-Think, instruction following drops as instructions accumulate, Hindi (24.5) and Russian (57.6) fall well below other languages, and every model stays below 5 on the τ³-bench banking subset.