BotPsy brings agentic tool use to small 0.8B models

Evidence level: Model benchmarking and developer-reported prototype evaluation

Why it matters: Running capable AI agents often requires substantial cloud infrastructure, which can increase cost, create privacy and data-control concerns, and make local execution on edge devices difficult. A small model designed for function calling and agentic workflows could make routine personal-assistant tasks more practical on consumer hardware.

How BotPsy processes instructions and tool calls BotPsy is a 0.8-billion-parameter model built on Qwen3.5-0.8B and designed for on-device deployment. It is intended to act as a personal assistant that uses tools for tasks including web search and research, travel and lifestyle, commerce and finance, business and customer operations, device and vehicle control, files and documents, calendar, email and messaging.

The model was post-trained in four sequential stages: supervised fine-tuning (SFT), multi-teacher on-policy distillation (MOPD) and two rounds of reinforcement learning. The first reinforcement-learning stage focused on tool calling, instruction following and STEM multiple-choice questions, using DAPO-based training with rule-based and verifier-derived rewards.

The first stage used supervised fine-tuning on multi-turn tool-use trajectories, including single, parallel and multi-step tool calls as well as cases in which no tool should be called. The distillation stage used specialist teachers—three RL-trained Qwen3.5-4B models—to recover knowledge and instruction-following performance that had degraded during SFT.

How BotPsy compares with larger models BotPsy roughly doubles the tool-use performance of its Qwen3.5-0.8B base model. Its Berkeley Function Calling Leaderboard version 4 (BFCL v4) score rises from 23.4 to 44.0; ACEBench rises from 30.1 to 58.7; and tau-bench rises from 23.2 to 51.5.

It also outperforms Qwen3.5-2B-Think—a model 2.5 times its size—on five of six agentic benchmarks: BFCL v4, ACEBench, tau-bench, ToolSandbox and BrowseComp-Plus. The exception is Claw-Eval, where Qwen3.5-2B-Think scores 48.9 versus BotPsy’s 46.5.

BotPsy is the strongest of the five models tested at or below 1.2 billion parameters. It also improves on its base model in general-knowledge benchmarks, reaching 38.1 on Graduate-Level Google-Proof Q&A (GPQA) Diamond and 48.9 on Massive Multitask Language Understanding Pro (MMLU-Pro), compared with 33.2 and 41.6 for the base model.

On GPQA Diamond and MMLU-Pro questions answered correctly by all five evaluated models, BotPsy used 3.1 times and 2.7 times fewer reasoning tokens, respectively, than its base model. The comparison was restricted to tasks solved by all five models, so differences in task difficulty do not explain the token reduction.

Safety and reliability under testing The available version of the Hugging Face post states that BotPsy still makes mistakes and is not intended for safety-critical or high-stakes automated decisions, or applications involving money, personal information, health or other sensitive topics.

The draft’s red-teaming figures—21% versus 36% unsafe responses, 46% versus 31% unsafe responses under prompt injection, 22.9% versus 1.4% repetitive-reasoning loops and 28.3% versus 0.2% runs reaching a 32,000-token limit—could not be verified from the page content retrieved for this review. They should not be presented as confirmed findings unless checked against the full evaluation section and its definitions of “unsafe response,” prompt-injection attack and failure mode.

An analysis of BotPsy’s performance and boundaries BotPsy’s gains come from post-training rather than from a larger parameter count. It uses multi-stage supervised fine-tuning, distillation from specialist teacher models and reinforcement learning with verifiable rewards to improve tool use, instruction following and reasoning efficiency.

The model is particularly strong in multi-turn, stateful interactions. On BFCL v4 multi-turn, its score rises from 7.5 to 46.1; on ACEBench Agent, from 7.9 to 28.3; and on tau-squared-bench, it reaches 67.8 in retail and 82.5 in telecom scenarios.

Its main limitations are open-ended search and some complex tasks. BrowseComp-Plus reaches only 4.5, although that is the best score among the five tested models; BFCL web search is 14.0 and memory is 16.3, both below Qwen3.5-2B-Think.

BotPsy was primarily trained and tested in English. It performs well in several additional languages but is substantially weaker in Hindi (24.5) and Russian (57.6), which remain notable limitations for multilingual deployment.

Availability and deployment format The announcement provides model downloads in Hugging Face safetensors format and GGUF files in bf16, Q8 and Q4 precisions. The available page content does not specify RAM, VRAM, battery consumption or device requirements for local deployment.

The post also does not establish that BotPsy is ready for autonomous execution of financial transactions, health-related decisions or other high-stakes workflows. Its developers explicitly position it as best suited to low- to medium-difficulty tasks and warn against safety-critical use.

What the benchmark scores do not show It may seem that BotPsy’s benchmark results mean it can safely automate personal tasks involving money or sensitive data. In fact: the developers state that it still makes mistakes and is not intended for safety-critical or high-stakes automated decisions, or applications involving money, personal information, health or other sensitive topics.

Strong relative gains do not establish reliable performance in production: BrowseComp-Plus remains low in absolute terms, and performance declines as multi-turn instruction complexity increases.

The comparison is developer-reported and benchmark-based; it does not demonstrate real-world reliability, security, privacy, battery life or performance across diverse consumer devices.

The claimed on-device suitability does not establish that a given phone, laptop or embedded device can run the model at useful speed or within acceptable memory and power limits.

Verdict: BotPsy demonstrates that a 0.8-billion-parameter model can substantially improve tool-use and agentic benchmark performance through multi-stage post-training, roughly doubling tool-use performance relative to its base model and outperforming a 2-billion-parameter comparison model on five of six agentic benchmarks. However, its open-ended search remains weak, multilingual performance is uneven, hardware requirements are unspecified, and the developers explicitly exclude safety-critical or high-stakes deployment.

Based on reporting from Hugging Face Blog. This is an independent explainer, written in our own words with AI assistance; Hugging Face Blog has not reviewed or endorsed it. Read the original for the full details.