BotPsy

BotPsy is the 0.8-billion-parameter on-device agentic language model Tether AI Research (QVAC) published on 11 November 2026 for function calling and personal-assistant workflows. The official write-up is the Hugging Face blog post BotPsy: On-Device Personal Assistant with a Sub-Billion-Parameter Agentic Language Model, dated November 11, 2026. The weight repo is qvac/BotPsy-0.8B; the quantized files are qvac/BotPsy-0.8B-GGUFs. The listed author is Tether Data, S.A. de C.V., doing business as Tether AI Research.

According to the official blog, BotPsy is built on Qwen3.5-0.8B and is positioned as a “personal assistant” that uses tools for simple tasks in web search and research, travel and lifestyle, commerce and finance, business and customer operations, device and vehicle control, files and documents, calendar, and email and messaging. Because of its small parameter count, the authors say it is best suited to tasks of low to medium difficulty. Training has four sequential stages: full-parameter supervised fine-tuning on tool-use conversations, multi-teacher on-policy distillation (MOPD) with three RL-trained Qwen3.5-4B specialist teachers, and two rounds of reinforcement learning with DAPO. The tool-calling reward checks format and then gives partial credit for matching tool names, argument names and argument values. The blog says this recovers knowledge and instruction following that dropped during SFT, while still improving tool use.

The authors report these evaluation numbers (all are research-team results published in the blog; they are not independently reproduced here): BFCL v4 rises from 23.4 for the base to 44.0, ACEBench from 30.1 to 58.7, and τ-bench from 23.2 to 51.5. In the ≤1.2B class they tested, BotPsy leads all six tool-use and agentic benchmarks, and on five of the six it also beats Qwen3.5-2B-Think, a model 2.5× its size. Instruction following improves sharply: IFEval from 59.7 to 90.8 and IFBench from 22.0 to 55.9. On BFCL v4’s hallucination tests, which check that the model declines when no appropriate tool exists, they report 86.1, the highest of the models tested. On efficiency, the blog says that on questions all five models answer correctly, BotPsy reaches the answer with about 3× fewer reasoning tokens than its base on GPQA Diamond and about 2.7× fewer on MMLU-Pro, and gets stuck in repetitive reasoning loops less often.

The same post lists the limits. Open-ended search remains weak: BrowseComp-Plus is 4.5, the best of the five but low in absolute terms. Multi-step information retrieval and research is the clear weak spot. The model was primarily trained and tested in English. On Multi-IF, Hindi (24.5) and Russian (57.6) lag well behind the other evaluated languages. BotPsy is a small model and it makes mistakes. The authors say it is not intended for safety-critical or high-stakes automated decisions, or for applications involving money, personal information, health or other sensitive topics. Weights are released under Apache 2.0 for research and educational use, in bf16, Q8 and Q4, as Hugging Face safetensors and as GGUF files for the QVAC SDK and llama.cpp. One sentence in the blog still says the model “will be released,” but the same page already links the model and GGUF repos, which are publicly reachable; this entry treats the release as published and defers to the repo pages.

Key links:

• Official blog: https://huggingface.co/blog/qvac/botpsy ; model: https://huggingface.co/qvac/BotPsy-0.8B ; GGUF: https://huggingface.co/qvac/BotPsy-0.8B-GGUFs ;
• The Hugging Face repos were created at 2026-11-10T03:10Z; the official blog is dated 2026-11-11. Earlier slots in this timezone today (00:32 through 16:32) missed the name; this slot records it ;
• This is a specific model name from Tether AI Research / QVAC. Do not merge it with the already-listed QVAC TranscribePsy, or with other models that use a Psy suffix.

Why it matters

This is a named on-device tool-calling model with an official blog, downloadable weights and a stated licence — not a vague fine-tune drop. The authors position it as a personal assistant small enough to run on a phone or laptop, and they publish comparisons against same-size and larger base models. For anyone tracking on-device agents and sub-billion-parameter tool use, that is a new, citable name.

The benchmark scores, reasoning-token figures and safety rates all come from the 11 November 2026 official blog and should be cited as research-team claims. Earlier scans in this timezone today, from 00:32 through 16:32, did not record the term; this slot adds it against the official blog date.

Limits

The official post limits intended use: low- to medium-difficulty tasks, primarily English, and not for high-stakes or sensitive automated decisions. Open-ended deep retrieval remains weak. The Apache 2.0 note in the blog is for research and educational use; the licence text is authoritative. Do not treat BotPsy as a name for the whole Tether or QVAC platform.

1. Read the official blog https://huggingface.co/blog/qvac/botpsy , and cross-check the model page https://huggingface.co/qvac/BotPsy-0.8B and the GGUF page https://huggingface.co/qvac/BotPsy-0.8B-GGUFs ;
2. Load the safetensors build or a bf16/Q8/Q4 GGUF in a local stack the official post names, such as the QVAC SDK or llama.cpp, and start with low-risk, non-sensitive tasks ;
3. When citing BFCL v4, ACEBench, τ-bench, IFEval, or the “decline when no tool fits” score, say the source is the research team’s results in the 11 November 2026 official blog, and re-check the upstream source ;
4. Do not treat BotPsy as QVAC TranscribePsy or as the entire Tether/QVAC platform; follow the official limits and do not use it for safety-critical decisions or for applications involving money, personal information or health; this page only restates the official blog and model pages — re-check the upstream source before citing it.