Cover: AI-generated editorial illustration.
On November 11, 2026, Tether AI Research published BotPsy, a 0.8-billion-parameter model fine-tuned from Qwen3.5-0.8B for tool use and multi-step agentic tasks. It is released under Apache 2.0 as Hugging Face safetensors and GGUF files (bf16, Q8, Q4), making it straightforward to run with llama.cpp.
What it can do
BotPsy is designed for tasks of low-to-medium difficulty: web search, calendar management, file operations, travel queries, and basic device control. The team is explicit that it is not intended for high-stakes decisions. Its benchmark scores, as reported by Tether, tell a clear story about where it earns its keep:
- BFCL v4: 44.0 vs. 23.4 for its base model, and 37.7 for Qwen3.5-2B-Think (a model 2.5x larger)
- Multi-turn τ-bench telecom: 82.5 vs. 69.3 for Qwen3.5-2B-Think
- Hallucination avoidance (knowing when not to call a tool): 86.1, the highest of all models tested
- Reasoning loop rate dropped from 22.9% to 1.4%, and token-limit hits from 28.3% to 0.2%
Training approach
The model went through four sequential stages: supervised fine-tuning on tool-use trajectories, multi-teacher on-policy distillation using three specialist Qwen3.5-4B models, then two rounds of reinforcement learning with rule-based rewards. Each stage built on the last, recovering general knowledge that SFT degraded and then pushing further on tool accuracy.
Why this matters practically
Running a capable agent locally removes per-call API costs, latency variance, and the risk of a provider deprecating access. For developers building automations where the task is well-defined and data privacy matters, a model this size can run on a laptop or mobile device. The weak spots are worth noting: open-ended web search (BFCL web search: 14.0), low-resource languages like Hindi, and complex banking workflows where every model tested scored below 5. Those cases still benefit from a larger cloud model.
Instruction following is a genuine strength: IFEval 90.8 and IFBench 55.9 are the highest of all models at 1.2B parameters or fewer tested by the team. If your agent needs to reliably follow structured prompts across multiple turns, that is a meaningful signal.