Playbooks · Coding tasks
BotPsy 0.8B: an agent model that fits on-device
Written November 11, 2026
Tether's BotPsy 0.8B leads every tested sub-1.2B model on six agent benchmarks — but 31% of prompt-injection replies are still unsafe.
What it is
Tether AI Research introduces BotPsy, a 0.8B-parameter model designed for function calling and agentic workflows, built on top of Qwen3.5-0.8B and suitable for on-device deployment. It is a personal assistant that uses tools to perform simple tasks for the user, in domains such as 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. Due to its small parameter count, it is best suited to tasks of low to medium difficulty. The post is plain about that limit: BotPsy still makes mistakes and is not intended for high-stakes use.
The training ran in four sequential stages: supervised fine-tuning, multi-teacher on-policy distillation, and two rounds of reinforcement learning. Three specialist Qwen3.5-4B teachers, each trained with reinforcement learning for one skill, guided the student during the distillation stage.
The benchmarks
BotPsy roughly doubles the agentic capability of its base model. BFCL v4 rises from 23.4 to 44.0, ACEBench from 30.1 to 58.7, and τ-bench from 23.2 to 51.5. It outperforms every model of up to 1.2B parameters tested, on all six tool-use and agentic benchmarks: BFCL v4, ACEBench, τ-bench, ToolSandbox, Claw-Eval, and BrowseComp-Plus. It also beats Qwen3.5-2B-Think, a model 2.5 times its size, on five of the six: BFCL v4 (44.0 vs 37.7), ACEBench (58.7 vs 55.4), τ-bench (51.5 vs 46.5), ToolSandbox (64.6 vs 41.2), and BrowseComp-Plus (4.5 vs 2.8). Only on Claw-Eval does Qwen3.5-2B-Think lead (48.9 vs 46.5).
Two numbers deserve a second look. BrowseComp-Plus at 4.5 is the best of the models tested, and it is low in absolute terms. And general knowledge still scales mainly with parameter count: GPQA Diamond rises from 33.2 to 38.1 and MMLU-Pro from 41.6 to 48.9, both best among the models at 1.2B parameters or below, but Qwen3.5-2B-Think scores well above on both (59.9 and 65.5). BotPsy is therefore designed to call tools for facts it does not know, not to serve as a knowledge model.
It is also more reliable in multi-turn work. On BFCL v4's multi-turn tasks it rises from 7.5 to 46.1 over its base, nearly double Qwen3.5-2B-Think (23.8). On BFCL v4's hallucination tests, which check that the model declines when no appropriate tool exists, BotPsy scores 86.1, the highest of all models tested.
The safety number
In automated red-teaming across six harm categories, BotPsy produces unsafe responses 21% of the time, compared with 36% for its base model. That is the lowest rate of all models of up to 1.2B parameters tested. Under prompt-injection attacks, which matter most for agents that read untrusted tool outputs, the unsafe rate drops from 46% for its base model to 31%. Better than its base, and better than its size class, but still unsafe in nearly a third of prompt-injection replies. For on-device use, where the model reads tool output you did not write, that is the limit that matters.
Getting it
The post links the weights as Hugging Face safetensors and as GGUF files for the QVAC SDK and llama.cpp, in three precisions (bf16, Q8, Q4), and states they are open under the Apache 2.0 license for research and educational use. Shorter, more decisive reasoning should reduce latency on-device: on questions all five models answer correctly, BotPsy reaches the answer with about three times fewer reasoning tokens than its base model on GPQA Diamond, and its rate of getting stuck in repetitive reasoning loops falls from 22.9% to 1.4%.
Limits
- The benchmarks are the authors' own (Tether AI Research and QVAC), not an independent measurement.
- BotPsy is best suited to tasks of low to medium difficulty and is not intended for high-stakes use.
- Prompt-injection replies are unsafe 31% of the time. Do not let it act on untrusted tool output without a check.
- BrowseComp-Plus at 4.5 is low in absolute terms, even though it leads the models tested.
Related playbooks
- Run Liquid AI's large 24-billion-parameter agent model on your own computer — if your memory fits it
- FlashProve: proofs of which model answered you
- SWE-2 Flash: Devin's cheaper model, closed weights
- DeepSeek, Exa, Brave, Tavily: which web-search API should your agent actually pay for now that search is metered
Sources
- Hugging Face Blog, BotPsy: On-Device Personal Assistant with a Sub-Billion-Parameter Agentic Language Model — the benchmark figures, the safety numbers, and the weights.
Back to: Coding playbooks hub · Also: TwIL-8B: a local model for math and coding · Addy Osmani's 25 agent skills for coding agents