BotPsy: a small on-device AI agent model built to act

QVAC BotPsy, released by Tether AI Research* on 11 November 2026, is a 0.8B-parameter model built to take actions for you, like replying to an email or adding a trip to your calendar. It is small enough to run on your own phone or laptop, so the decisions about your inbox and calendar are made on the device. The weights are free to download on Hugging Face under Apache 2.0.

Why should an AI agent run on your device?

AI assistants are starting to do things for us: book a trip, sort an inbox, send a reminder, rebuild a calendar when plans change. To do that well, an assistant needs the most personal context we have, from emails and contacts to travel plans and accounts. Today that usually means sending all of it to a few companies' servers and trusting them with it.

We think you should not have to choose between a capable assistant and your privacy. An assistant that runs on your own device can read your inbox and change your calendar without the details of your day reaching a data center. A model on the device also keeps working without a connection, and nobody bills you per request.

Until now, size was the main obstacle. A phone gives each app a small slice of memory, and the models small enough to fit in it made too many mistakes to be trusted with actions: they called the wrong tool, or invented one that does not exist. BotPsy starts from one of those small models, Qwen3.5-0.8B, and was trained in four stages to make it more reliable at calling tools.

What can BotPsy do?

BotPsy acts through tool calls. An app gives it a list of tools, which are functions such as get_email, send_email or create_event. When you ask "add the dentist on Friday at 3 pm", BotPsy picks the relevant function and fills in the details: create_event(title="Dentist", day="Friday", time="15:00"). The app runs that tool call and sends back the result. BotPsy reads it and carries on until the task is done.

BotPsy is built for everyday requests in six groups of apps:

• Your phone and computer: alarms and timers, files and documents, device settings.
• Work: calendar and tasks, email, collaboration tools.
• Messages: sending and answering in chat apps.
• Finding things out: web search, weather, news.
• Daily life and media: music, video, photos, recipes and fitness.
• Errands: maps and travel, shopping and delivery, finance.

Use case example:

BotPsy and its base model got the same request in a test from the BFCL v4 benchmark: top up the car so the tank holds 50 gallons, start the engine and make sure the car is locked. It already had 10 gallons in it, so the right amount to add was 40.

• Qwen3.5-0.8B, the base model, tried to add 50 gallons and was told the tank cannot hold that much. It then called tools that do not exist and reported the car ready. The tank was never filled.
• BotPsy checked the tank first, found 10 gallons and added 40. When the engine refused to start, it locked the doors, pressed the brake and started it.

On single requests, such as "set an alarm for 7" or "what is the weather in Lisbon tomorrow", BotPsy scores between 69 and 79 out of 100 in every group of apps with enough test cases. Tasks that take several steps are harder for every model we tested. On those, BotPsy averages 50.3 out of 100 over seven groups of apps, the six above plus one for other and multi-app tasks. That is more than twice the base model's 22.1 and ahead of Qwen3.5-2B (42.5), a model 2.5 times its size.

Note: Each scenario was first assigned to an app group from the benchmark's own metadata where it names the app (for example the task domain, the API class or the tool service of the gold calls). The remaining scenarios were classified by an LLM.

What happens when a step fails?

Real tasks rarely go to plan, so an agent has to notice what went wrong and fix it.

In one ACEBench test, BotPsy had to send Frank a reminder about tomorrow's meeting. The first send failed because Wi-Fi was off, and the second because the inbox was full. The model BotPsy was built from kept retrying the same send, and the message never went out. BotPsy turned Wi-Fi on, deleted a message to make room, and sent it.

How good is it for its size?

On four agent benchmarks, BFCL v4, ACEBench, τ-bench and ToolSandbox, BotPsy roughly doubles its base model's scores and is also ahead of Qwen3.5-2B. Across all six tool-use and agent benchmarks, it leads every model of 1.2B parameters or fewer that we tested.

BotPsy gained most in long conversations where it has to keep track of a task. On the multi-turn part of BFCL v4, BotPsy went from 7.5 to 46.1, almost twice the 23.8 of Qwen3.5-2B. On τ²-bench telecom, where the agent guides a customer through troubleshooting over many turns while following a strict policy, it scores 82.5, against 69.3 for Qwen3.5-2B.

It also follows instructions closely: 90.8 on IFEval, up from 59.7. Its general knowledge improved too, but knowledge still grows with model size, so BotPsy is built to look up facts it does not know with a search tool rather than answer from memory.

Does it know when not to act?

When none of the available tools fits a request, BotPsy usually says so instead of inventing a call: it scores 86.1 on the BFCL v4 test for this, the highest of all the models tested.

It is also harder to push into a harmful answer. In automated attack tests across six kinds of harm, it answered unsafely 21% of the time, the lowest rate among the models of 1.2B parameters or fewer that we tested, against 36% for its base model. Against prompt injection, where an attacker hides instructions in an email or a web page the agent reads, the unsafe rate fell from 46% to 31%. For apps with strict safety needs, we recommend adding a separate filter on what goes in and what comes out.

Why does it answer faster?

BotPsy answers faster than its base model because it reaches the right answer with much less thinking. Reasoning tokens are the notes a model writes to itself before answering. On GPQA Diamond, a set of hard science questions, BotPsy writes 3,690 of them per correct answer against 11,563 for its base, 68% fewer, on the questions that every model in the comparison answered correctly, and it scores higher. It also stopped going in circles: on IFBench, the share of answers stuck repeating the same reasoning fell from 22.9% to 1.4%. In long customer conversations it needs 9 to 26% fewer turns than its base to finish the job.

Where is BotPsy going?

BotPsy points toward personal AI that will live on your own devices, understand what you want, act through the apps you already use, and keep your data, passwords and decisions on your hardware.

QVAC Fabric will let you fine-tune the model on your own hardware until it fits your tools and habits. Web searches will be able to go through a peer-to-peer network Tether AI Research is building, over an encrypted link, so your queries will not reach big centralized search providers.

Your phone will be the interface you talk to. When a job is too heavy for it, such as analyzing a long document or untangling a messy travel plan, the work will go to your own computer at home or to trusted peer-to-peer nodes, and the result will come back to your phone. In that setup you will own the whole system, and your data will stay yours.

How do I run it?

BotPsy is on Hugging Face in two precision variants, BF16 and 8-bit, as standard weights and as GGUF files for llama.cpp. It keeps the architecture of Qwen3.5, so llama.cpp and the Hugging Face libraries run it without custom code.

It works best in English, the language it was trained and tested in. On a multilingual instruction test it scores 72 to 80 in English, Chinese, French, Spanish, Italian and Portuguese, lower in Russian and much lower in Hindi. It is built for tasks of low to medium difficulty: questions that take many web searches in a row remain hard for a model this size.

BotPsy is a small language model developed to advance the art and support research and educational uses. It makes mistakes. It is not intended for safety-critical or high-stakes automated decisions, applications involving money, personal information, health and other sensitive topics.

FAQ

Get BotPsy

• Models on Hugging Face
• The full write-up with every benchmark and the training method

Sources

• Tether AI Research, QVAC BotPsy article and model cards on Hugging Face, 11 November 2026.
• Qwen Team, Qwen3.5-0.8B
• Berkeley Function Calling Leaderboard (BFCL v4)
• Chen et al., ACEBench
• Yao et al., τ-bench
• Barres et al., τ²-bench
• Lu et al., ToolSandbox
• Zhou et al., IFEval
• Pyatkin et al., IFBench
• Rein et al., GPQA
• Confident AI, DeepTeam red-teaming framework

* References to Tether AI Research mean Tether Data, S.A. de C.V.

Updated 11 November 2026.