BotPsy: The 0.8B On-Device Agent Model from Tether AI Research, Explained

BotPsy is a 0.8B-parameter language model from Tether AI Research, released on November 11, 2026, that is built specifically for function calling and agentic workflows on-device. It is post-trained on top of Qwen3.5-0.8B, published with open weights under the Apache 2.0 license, and aimed at the kind of personal-assistant jobs that today usually go to a cloud API: web search, calendar and email, files and documents, device control, travel and shopping. The pitch is that a model small enough to run on a laptop or phone can now do routine agent work privately, with no API keys and no per-call cost. Source

Why a sub-billion-parameter agent matters

Most agent deployments today call a frontier model in the cloud, and Tether AI Research's framing of the problem will sound familiar to anyone who has run one: costs that fluctuate, providers that can modify or deprecate access, limited customization, and user data leaving the device. Their argument is that many routine workloads do not need frontier-level intelligence, and that a small, specialized model can handle them cheaper and privately. That is the same reasoning behind the cost-control work we see teams do once an agent goes to production, where routing simple tasks to smaller models is often the single biggest lever on the bill. Source

BotPsy is aimed squarely at that gap. It is a "personal assistant" style model that uses tools across domains like 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. Tether is explicit that, at 0.8B parameters, it is best suited to tasks of low to medium difficulty. Source

What the benchmarks show

The headline numbers come from Tether AI Research's own evaluation, run on November 3, 2026, comparing BotPsy against its Qwen3.5-0.8B-Think base and three small rivals: Qwen3.5-2B-Think (2.5 times its size), LFM2.5-1.2B-Think and MiniCPM5-1B-Think. All models were evaluated in thinking mode, and all scores are percentages. These are vendor-reported figures, so treat them as the lab's own measurements rather than independent results. Source

On tool use and agentic benchmarks, BotPsy roughly doubles its base model and leads five of six tests. BFCL v4 rises from 23.4 to 44.0, ACEBench from 30.1 to 58.7, tau-bench from 23.2 to 51.5, ToolSandbox from 30.5 to 64.6, and BrowseComp-Plus from 1.9 to 4.5. On all five it also beats Qwen3.5-2B-Think, which scores 37.7, 55.4, 46.5, 41.2 and 2.8 respectively. The one exception is Claw-Eval, where the 2B model stays ahead at 48.9 to BotPsy's 46.5. Source

The largest gains are in multi-turn, stateful work, which is where agents actually live. On BFCL v4's multi-turn tasks BotPsy jumps from 7.5 to 46.1, nearly double Qwen3.5-2B-Think's 23.8. On tau-2-bench telecom, a long troubleshooting conversation where the agent must follow policy and act for the customer, it scores 82.5 against 69.3 for the 2B model. The one multi-turn test where it does not lead is tau-3-bench banking, where no model gets above 5. Source

Two quieter results matter as much for production use. First, BotPsy is the most reliable of the tested models at not calling a tool when none fits, scoring 86.1 on BFCL v4's hallucination tests. Second, instruction following improves sharply: IFEval rises from 59.7 to 90.8 and IFBench from 22.0 to 55.9, the best in its class, though its Multi-IF scores show a real weakness in Hindi (24.5) and Russian (57.6) compared with 72 to 80 in English, Chinese, French, Spanish, Italian and Portuguese. Source

Safety and efficiency

For an agent that reads untrusted tool output, injection resistance is not optional. In Tether's automated red-teaming across six harm categories, BotPsy produced unsafe responses 21 percent of the time, down from 36 percent for its base model and the lowest rate among models up to 1.2B parameters tested. Under prompt-injection attacks specifically, the unsafe rate drops from 46 percent to 31 percent. That is an improvement, not a pass: a 31 percent failure rate under attack is still a number that argues for the kind of least-privilege tool design and output filtering we recommend for any agent, whatever its size. Source

On efficiency, Tether reports that BotPsy reaches correct answers with 1.4 to 3.1 times fewer reasoning tokens than its base on GPQA Diamond, MMLU-Pro, IFEval and IFBench, measured on questions all five models answered correctly. Its rate of getting stuck in repetitive reasoning loops on IFBench falls from 22.9 percent to 1.4 percent, and runs hitting the 32k-token limit fall from 28.3 percent to 0.2 percent. Shorter, more decisive reasoning translates directly into lower latency on-device. Source

Where it is weak

Tether is unusually candid about the weak spots, and they are worth quoting directly. Open-ended search is the big one: BrowseComp-Plus at 4.5 is the best in class but low in absolute terms, and on BFCL's web search (14.0) and memory (16.3) subsets BotPsy trails Qwen3.5-2B-Think. Complex information retrieval scenarios average just 15, so deep multi-step research is not this model's job. The model card and blog both stress that BotPsy makes mistakes, is not intended for safety-critical or high-stakes decisions, and should not be used for applications involving money, personal information, health or other sensitive topics. English is the recommended language. Source

How it was trained and how to run it

The training recipe has four stages: supervised fine-tuning, multi-teacher on-policy distillation guided by three specialist Qwen3.5-4B teachers, and two rounds of reinforcement learning with rule-based tool-calling rewards. The weights are the interesting part for practitioners. BotPsy is released under Apache 2.0 for research and educational use, as BF16 safetensors in the main repository and as GGUF files in BF16, Q8_0 and Q4_K_M in a companion repository, ready for llama.cpp, the QVAC SDK, Ollama, LM Studio or any compatible app. Source

The QVAC SDK is Tether's open-source local AI runtime: GGUF support, RAG, speech-to-text and P2P inference, cross-platform across Linux, macOS, Windows, Android and iOS, with no cloud and no API keys, under Apache 2.0. For teams building on-device agent hardware, from voice assistants to ESP32-class edge devices, a 0.8B tool-calling model in Q4 is the kind of artifact that makes local-first designs realistic, and it slots into the same privacy argument we made when covering on-device agent security for smart gadgets. Source

Our read

This is our assessment, not Tether's. BotPsy is worth attention for three reasons. First, the multi-turn numbers, if they hold up under independent testing, mark a real capability step at this size, and multi-turn reliability is exactly what separates a demo agent from a useful one. Second, the explicit "knows when not to call a tool" training shows in the hallucination score, and that trait is underrated in agent design. Third, the honest limitations section makes the model easier to adopt responsibly. The caveats are equally real: every number here is self-reported, the knowledge benchmarks confirm this is a tool-user rather than a knowledge model, and the weak search and memory scores mean a production assistant still needs a retrieval layer or a bigger model behind it. If you are building a local-first assistant, it is a strong candidate for the tool-calling tier, paired with guardrails and a clear escalation path. If you are weighing cloud agents against on-device ones, our earlier pieces on AI agent cost control and ESP32 AI voice agents cover the surrounding trade-offs, and our look at agent device security for smart gadgets covers the threat model that comes with putting agents on hardware.