
BotPsy: A 0.8B-Parameter On-Device Agentic Language Model

Tether AI Research introduces BotPsy, a 0.8B-parameter model built on Qwen3.5-0.8B for function calling and agentic workflows on-device. The motivation is cost, control, and privacy: frontier cloud LLMs have fluctuating operational costs, can be modified or deprecated by providers, limit customization, and expose user data. Not every task needs frontier-level intelligence, and smaller specialized agents can be equally capable on specific tasks at much lower cost. BotPsy is positioned as a personal assistant that uses tools for domains including web search, travel, commerce, finance, business operations, device control, files, calendar, email, and messaging.
The model is post-trained in four sequential stages: supervised fine-tuning (SFT) on tool-use conversations, multi-teacher on-policy distillation (MOPD) using three RL-trained Qwen3.5-4B specialist teachers, and two rounds of reinforcement learning with DAPO. Rewards are rule-based and verifier-based; the tool-calling reward follows Tool-N1’s format checking but gives fine-grained partial credit in the spirit of ToolRL, matching predicted tool names, argument names, and argument values against ground truth. Each stage uses a different data mixture, with SFT focused on tool use, MOPD recovering knowledge and instruction-following regressions from SFT, and the second RL stage shifting toward general-domain verifiable tasks.
Results are reported against four comparison models up to 1.2B parameters plus Qwen3.5-2B-Think at 2.5× BotPsy’s size. BotPsy roughly doubles its base model on key agentic benchmarks: BFCL v4 rises from 23.4 to 44.0, ACEBench from 30.1 to 58.7, and τ-bench from 23.2 to 51.5. It leads all six tool-use benchmarks in the ≤1.2B class and beats Qwen3.5-2B-Think on five of six, including BFCL v4 (44.0 vs 37.7), ACEBench (58.7 vs 55.4), τ-bench (51.5 vs 46.5), ToolSandbox (64.6 vs 41.2), and BrowseComp-Plus (4.5 vs 2.8). Only on Claw-Eval does the larger model lead (48.9 vs 46.5).
Multi-turn behavior improves sharply. On τ²-bench telecom it scores 82.5 versus 69.3 for Qwen3.5-2B-Think, and on BFCL v4 multi-turn it rises from 7.5 to 46.1. Tool-calling restraint also improves: on BFCL v4 hallucination tests BotPsy scores 86.1, the highest tested, and on ACEBench special cases it rises from 13.7 to 51.3. Instruction following improves from IFEval 59.7 to 90.8 and IFBench 22.0 to 55.9, the highest IFBench score tested. General knowledge is preserved or improved: GPQA Diamond rises from 33.2 to 38.1 and MMLU-Pro from 41.6 to 48.9, though knowledge still scales mainly with parameter count and BotPsy is designed to call tools for facts it does not know.
Safety and efficiency also improve. In automated red-teaming across six harm categories, BotPsy produces unsafe responses 21% of the time versus 36% for its base, the lowest among models up to 1.2B parameters. Under prompt injection, the unsafe rate drops from 46% to 31%. BotPsy reaches correct answers with roughly 3× fewer reasoning tokens on GPQA Diamond (3,690 vs 11,563) and 2.7× fewer on MMLU-Pro (1,912 vs 5,105). Repetitive reasoning loops fall from 22.9% to 1.4% on IFBench, and runs hitting the 32k-token limit drop from 28.3% to 0.2%. On τ²-bench solved tasks it needs 26% fewer assistant turns than the base in telecom.
The source also states explicit limits. BotPsy is best suited to low- or medium-difficulty tasks. Multi-step information retrieval is the weakest area: complex Information Retrieval & Research scenarios score 15, mostly BrowseComp-Plus deep-research questions, which remain hard for every model in this size range. Lower-resource languages lag, with Hindi at 24.5 and Russian at 57.6 on Multi-IF. Performance drops as instructions accumulate across turns. The model is not intended for safety-critical or high-stakes decisions, or for applications involving money, personal information, health, or other sensitive topics. It was primarily trained and tested in English. BotPsy is released with open weights under Apache 2.0 for research and educational use, in bf16, Q8, and Q4 precisions, as Hugging Face safetensors and GGUF files for the QVAC SDK and llama.cpp.

