The development
What happened
Tether AI Research has introduced BotPsy, a 0.8-billion-parameter language model finetuned from Qwen3.5-0.8B for function calling and agent workflows. The 11 November research announcement targets low-to-medium-difficulty tasks; its model card explicitly excludes high-stakes decisions and applications involving money, personal information, health or other sensitive topics.
Key facts
The model card lists BF16 safetensors and linked BF16, Q8_0 and Q4_K_M GGUF files under an Apache 2.0 declaration. Training code and training data are not included in the release.
Post-training uses supervised fine-tuning, multi-teacher on-policy distillation and two rounds of reinforcement learning, according to the developers.
Tether reports a BFCL v4 weighted overall score of 44.0, versus 23.4 for its base model, averaged over three runs. These are developer evaluations, not independent task-success guarantees.
The model-card comparisons used thinking mode and a shared harness, with BotPsy served as BF16 through vLLM. They do not establish phone latency or quantized-model performance.
Open-ended search remains weak: BrowseComp-Plus scores 4.5 out of 100. English is recommended; the authors report weaker Hindi and Russian results and degradation as instructions accumulate.
The blog’s automated red-team tests still produced unsafe responses in 21% of cases and 31% under prompt injection. Lower rates than its base model are not a safety or privacy certification.
The context
Why it matters
TideReach assessment: small, downloadable tool-use models offer a research path toward local assistants without making frontier-model capability or end-to-end privacy automatic. BotPsy’s server-side benchmark gains, remaining unsafe outputs and explicit sensitive-use exclusions make bounded evaluation and tool permissions more important than the on-device label.