Tether AI Research announced BotPsy on November 11, 2026. The small AI model is intended for developers building applications that carry out simple tasks on a user’s device—not people looking for an assistant they can install and immediately use.
For small teams exploring local assistants, the useful decision is whether a narrowly defined task fits the model’s limits, and whether a smaller download is worth the reported quality tradeoff.
What you get—and what you still have to build
The model card describes a text-only model with about 0.8 billion parameters. Tool calling means turning a request into a structured instruction for another piece of software. BotPsy proposes tool calls; the surrounding application must provide the tools and control their execution.
The public model repository lists downloadable weights under the Apache 2.0 license. The QVAC integration guide describes building an application around local models, including native mobile builds. Its mobile instructions require a physical device; emulators are currently unsupported.
If you want an existing phone feature instead, see the guide to Gemini’s Android task automation. That is a separate product with its own access and data-handling rules, not an equivalent local-model option.
Compare download size and reported quality
GGUF is a model-file format used by local inference software. BotPsy’s repository lists three versions; the developer recommends Q8_0 and publishes the following scores. The final column is Tether’s average across ten benchmarks, not an independently verified success rate for an app.
Version | Model-file size | Vendor’s ten-benchmark mean |
|---|---|---|
BF16, unquantized | 1.52 GB | 50.4 |
Q8_0, eight-bit | 0.81 GB | 50.7 |
Q4_K_M, four-bit | 0.53 GB | 48.9 |
Calculation: using the repository’s exact byte counts, choosing Q4 instead of Q8 saves 282,546,080 bytes—about 283 MB, or 34.8%. Its reported average is 1.8 points lower than Q8’s. The storage saving is not an equivalent reduction in running memory, and the score gap is not an estimate of errors in your application.
Sizes here use decimal GB and describe model files, not total memory requirements. Tether ran the GGUF evaluations on NVIDIA H100 GPUs using llama.cpp. Those results do not establish phone response times, battery use or sustained performance.
Check the task limits before choosing a file
Tether’s intended-use guidance describes BotPsy as supporting research and education, recommends English, and says it is best suited to low-to-medium difficulty tasks. It says the model is not intended for applications involving money, personal information, health or other sensitive topics, or for high-stakes automated decisions. Running locally does not remove those limits.
The announcement’s scenario breakdown reports an overall score of 75 for simple, single-call tasks and 39.5 for complex tasks. These are vendor results on different task sets—not evidence that adding one step halves an app’s reliability. They support starting with a tightly bounded task, not assuming general personal-assistant competence.
There is also a practical distinction between local inference and an entirely local application. The model may run on the device while a connected tool sends a request elsewhere. Review the tools and logging as part of the integration; the model’s location alone cannot establish the app’s privacy properties.
How to make the first evaluation useful
A reasonable starting point is an English-language prototype with invented, non-sensitive inputs and one reversible demo action, such as changing a setting in a test app. Check the selected action and its arguments, including cases where the correct response is to ask for clarification or do nothing.
Use Q8 as the developer-recommended starting point if the target device has enough capacity. If download size makes Q4 attractive, compare both on the same tasks. Measure actual memory use and response time on that device before deciding; the file-size table cannot answer those questions.
One setup detail matters: the documented chat template defaults to non-thinking mode, while the published evaluations enabled thinking. Follow the documented enable_thinking setting when comparing with those results. A successful local load is only the start—the decision should rest on whether the complete prototype handles its intended tasks acceptably.