A 0.8B on-device assistant beats a model 2.5 times its size on five of six tool-use benchmarks

A model under one billion parameters is trained specifically as a tool-calling personal assistant that runs locally on a device, covering everyday tasks such as web search, calendar, email and device control. The prior option was calling frontier cloud LLMs, which carry high and fluctuating operating costs, and whose providers can modify, deny or deprecate access at any time; not every task needs frontier-level intelligence. Across six tool-use and agentic benchmarks, BotPsy scores highest among the tested models on five: BFCL v4 44.0, ACEBench 58.7, τ-bench 51.5, ToolSandbox 64.6 and BrowseComp-Plus 4.5, all above the 2.5-times-larger Qwen3.5-2B-Think at 37.7, 55.4, 46.5, 41.2 and 2.8; only on Claw-Eval does Qwen3.5-2B-Think lead, 48.9 to 46.5.

Training proceeds in four stages: imitating tool-use conversations, learning from three specialist teachers, then two rounds of reinforcement learning scored by rules. Supervised fine-tuning raises tool use but pushes general knowledge and instruction following below the base model, and the later stages have to bring both back. The first stage is full-parameter supervised fine-tuning on tool-use conversations; the second is on-policy distillation in which the student generates its own responses and three Qwen3.5-4B teachers, each RL-trained for one skill and routed per prompt, supply a per-token reverse KL signal; the third and fourth stages use DAPO reinforcement learning with rewards computed by rules and verifiers, where the tool-calling reward gives fine-grained partial credit for predicted tool names, argument names and argument values.

The model is primarily trained and tested in English, which is the recommended language; it performs well in most of the seven additional languages covered by Multi-IF but scores much lower in Hindi and Russian. It is intended for research and educational use and is best suited to low-to-medium difficulty tasks; it is not intended for safety-critical or high-stakes automated decisions, or for applications involving money, personal information, health and other sensitive topics.

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. Instruction following drops as instructions accumulate, reaching 51.0 in turn three versus 56.5 for Qwen3.5-2B-Think and 56.2 for LFM2.5-1.2B-Think; on the τ³-bench banking subset every model stays below 5.