📅 November 11, 2026 • Tether AI Research
Tether AI Research introduced BotPsy, a 0.8B-parameter model designed for function calling and agentic workflows, built on top of Qwen3.5-0.8B and suited for on-device deployment. The announcement was published on the Hugging Face blog on November 11, 2026. Core benchmarks roughly double the base model: BFCL v4 rises from 23.4 to 44.0, ACEBench from 30.1 to 58.7, and tau-bench from 23.2 to 51.5 (mean of the retail, telecom and banking subsets).
Per Tether AI Research, BotPsy outperforms every tested model up to 1.2B parameters — Qwen3.5-0.8B-Think, LFM2.5-1.2B-Think and MiniCPM5-1B-Think — on six tool-use and agentic benchmarks: BFCL v4, ACEBench, tau-bench, ToolSandbox, Claw-Eval and BrowseComp-Plus. It also beats the 2.5x larger Qwen3.5-2B-Think on 5 of 6 benchmarks, with Claw-Eval the only one where the larger sibling stays ahead. Instruction following rises from IFEval 59.7 to 90.8 and IFBench 22.0 to 55.9, and general knowledge improves too: GPQA Diamond from 33.2 to 38.1 and MMLU-Pro from 41.6 to 48.9.
Safer answers, fewer loops, sub-1 GB files
Safety was probed with automated red-teaming across six harm categories: BotPsy produces unsafe responses 21 percent of the time versus 36 percent for its base model, and under prompt-injection attacks the unsafe rate drops from 46 percent to 31 percent. Efficiency: on questions all five models answer correctly it needs about 3x fewer reasoning tokens on GPQA Diamond (3,690 vs 11,563), its rate of getting stuck in repetitive reasoning loops falls from 22.9 percent to 1.4 percent, and 32k-token-limit runs fall from 28.3 percent to 0.2 percent.
Training ran in four sequential stages: supervised fine-tuning, multi-teacher on-policy distillation with three RL-trained Qwen3.5-4B teachers, and two rounds of reinforcement learning. Weights are open under Apache 2.0 for research and education, safetensors and GGUF on Hugging Face: BF16 about 1.52 GB, Q8_0 about 0.81 GB, Q4_K_M about 0.53 GB, for llama.cpp and the QVAC SDK. Limits: best on low-to-medium difficulty tasks, primarily English with weaker Hindi and Russian, and no safety-critical or high-stakes automated decisions or apps involving money, personal information or health.
Sources: huggingface.co/blog/qvac/botpsy — huggingface.co/qvac/BotPsy-0.8B — huggingface.co/qvac/BotPsy-0.8B-GGUFs
Continue the conversation: https://t.me/AIAutomationArchitecture · Bluesky
