Tether AI Research published details on November 11, 2026 about BotPsy, a 0.8B-parameter model derived from Qwen3.5-0.8B and intended for on-device deployment. The group describes BotPsy as a “personal assistant” optimized for function calling and agentic workflows across domains such as web search, travel, commerce, device control, and messaging. The team emphasizes the model is best suited to low- and medium-difficulty tasks and is not intended for high-stakes use.
Design and training
According to Tether AI Research, BotPsy was post-trained in four stages: Supervised Fine-Tuning (SFT), Multi-teacher On-Policy Distillation (MOPD) guided by three Qwen3.5-4B specialist teachers, and two rounds of Reinforcement Learning (RL). The training mix targeted tool use, instruction following and verifiable rewards; the RL stages used rule-based tool-calling rewards and environment-verifiable tasks.
The team reports the model preserves general knowledge while improving instruction following and tool use. BotPsy will be released with open weights under the Apache 2.0 license and provided in three precisions (bf16, Q8, Q4) as Hugging Face safetensors and GGUF files for the QVAC SDK and llama.cpp.
Benchmarks, safety and limits
Tether AI Research presents benchmark results showing large gains over its Qwen3.5-0.8B base: BFCL v4 rose from 23.4 to 44.0, ACEBench from 30.1 to 58.7, and τ-bench from 23.2 to 51.5. The team reports BotPsy outperformed models up to 1.2B parameters on five of six tool-use and agentic benchmarks and scored higher than Qwen3.5-2B-Think on five benchmarks; Qwen3.5-2B-Think led only on Claw-Eval.
On safety metrics, the researchers report BotPsy produced unsafe responses 21% of the time in automated red-teaming versus 36% for its base model, and its unsafe rate under prompt-injection attacks fell from 46% to 31%. Efficiency gains include fewer reasoning tokens—3,690 versus 11,563 on GPQA Diamond and 1,912 versus 5,105 on MMLU-Pro—and a reduction in repetitive reasoning loops from 22.9% to 1.4%.
Tether AI Research notes BotPsy is designed for private, on-device workflows and highlights remaining weaknesses in open-ended search and some multi-lingual instruction-following cases. The release packaging and license information are specified for research and educational use under Apache 2.0.
Original source: Hugging Face Blog