Tether's 0.8B-parameter BotPsy runs a tool-using assistant on your own device
Tether AI Research tuned a tiny model to use tools like email and calendar, and says it beats every rival of similar size it tested.
The short version
- BotPsy has 0.8B parameters and is built on Qwen3.5-0.8B, aimed at low to medium difficulty tasks.
- BFCL v4 tool-use score rose from 23.4 to 44.0 against its base model.
- It beat the 2.5× larger Qwen3.5-2B-Think on five of six agent benchmarks, but not Claw-Eval (46.5 vs 48.9).
- Unsafe replies in automated red-teaming fell from 36% to 21%; prompt-injection failures from 46% to 31%.
- Weights are planned under the Apache 2.0 license, for research and educational use.
Why it matters
A phone-sized model that can reliably pick and call tools makes a private, offline assistant more realistic, with no per-request fee. The catch is that all results come from the developer's own tests, and the model is explicitly not meant for high-stakes use.
A tiny model with a job to do
As huggingface.co reported, Tether AI Research has published BotPsy, a language model with 0.8B parameters. Parameters are the adjustable numbers inside a model, and this count is small next to the giant models behind most cloud chatbots. BotPsy starts from Qwen3.5-0.8B and is trained to be an agentic model, meaning it takes actions by calling tools instead of only writing text.
The intended role is a personal assistant. The listed areas include web research, travel, shopping and finance, customer operations, device and vehicle control, files, calendar, and email and messaging. Tether says the model suits low to medium difficulty tasks, and that it still makes mistakes.
The pitch is practical. Cloud models can get pricier without warning, can be changed or withdrawn by the provider, and send your data off your device. A small model that runs locally avoids those problems for routine work.
How it was trained
Training had four stages, each starting from the best result of the one before. The first was supervised fine-tuning, where the model imitates example conversations that use tools, including cases where the right move is to call nothing.
The second stage was distillation, where a student model learns from stronger teachers. Here the student wrote its own answers and teacher models scored every token. The teachers were larger Qwen3.5-4B models, each tuned for a different skill. This stage mainly won back general knowledge and instruction following that had slipped during the first stage.
The last two stages used reinforcement learning, where the model is rewarded for correct results checked by rules. The tool-calling reward gave partial credit for getting the tool name, argument names and argument values right, instead of simple pass or fail. The second round added math, code and multi-step tool use in a simulated workplace.
What the benchmarks say
Tether compared BotPsy with other models of up to 1.2B parameters, plus the larger Qwen3.5-2B-Think. Against its own base, tool-use scores roughly doubled: BFCL v4 went from 23.4 to 44.0, ACEBench from 30.1 to 58.7, and τ-bench from 23.2 to 51.5. It led all small rivals on all six agent benchmarks.
It also topped the 2.5× larger Qwen3.5-2B-Think on five of the six, for example ToolSandbox (64.6 vs 41.2). It lost on Claw-Eval, 46.5 to 48.9.
Two results matter for an assistant that acts for you. On BFCL v4's tests of whether a model declines when no suitable tool exists, it scored 86.1, the highest of all models tested. And in a telecom customer-support simulation, it scored 82.5, ahead of the 2B model's 69.3.
General skills improved too. IFEval, which checks following explicit instructions, climbed from 59.7 to 90.8, and MMLU-Pro, a knowledge test, from 41.6 to 48.9.
Safety, efficiency and weak spots
In automated red-teaming, where attackers probe the model for harmful output, BotPsy gave unsafe replies 21% of the time against 36% for its base. Under prompt injection, where malicious text hidden in a tool's output tries to hijack the agent, the unsafe rate fell from 46% to 31%. That is still nearly a third, so it is not a solved problem.
It also thinks more briefly. On one knowledge test it needed 3,690 reasoning tokens against 11,563 for its base, on questions both answered correctly. Runs stuck in repetitive reasoning loops on IFBench dropped from 22.9% to 1.4%. Less rambling should mean faster responses on a phone.
The weaknesses are clear. Open-ended search is poor: BrowseComp-Plus is 4.5, the best of the group but low in absolute terms. Knowledge still scales with size, and the 2B model scores far higher on knowledge tests, so BotPsy is designed to look facts up with tools. Language coverage is uneven too, with Hindi at 24.5 on a multilingual instruction test and Russian at 57.6.
Availability and what comes next
Tether says it will release the model in three precisions, meaning how compactly the numbers are stored: bf16, Q8 and Q4. They come as standard Hugging Face files and as GGUF files, a format used by local-running tools such as llama.cpp and Tether's own QVAC SDK. The license is Apache 2.0, for research and educational use.
Tether's roadmap goes further: private peer-to-peer search, local fine-tuning to fit personal habits, handing heavy jobs from a phone to a home PC, and agents talking to each other on trusted networks. Those are plans, not features shown in the results.
What to watch
- Do independent testers reproduce Tether's benchmark numbers once the weights are public?
- How well does it run on real phones and laptops, since hardware needs are not covered here?
- Can prompt-injection failures fall well below 31% before people trust it with email and files?
Try it
Check the Hugging Face page to see whether the files are live. If so, run the Q4 version in llama.cpp on a few routine tasks, plus one request no tool can handle, and see whether it declines.
Sources and links
First reported by huggingface.co. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.