BotPsy brings tool calling to 0.8B, with a 530MB option
Tether announced BotPsy on November 11, a 0.8 billion parameter model specialized for turning requests into tool calls. Developers can download its weights rather than depend on a hosted service. The useful question is not whether it replaces a frontier assistant, but whether it can handle a narrow workflow on hardware you already own.
There is no BotPsy consumer app. The official product page describes a model for developers building assistants, with local inference and support for everyday tasks of low to medium difficulty. Training code and training data are not included in the Apache 2.0 open-weight release.
Choose the download, not just the parameter count
The official GGUF repository offers three versions for llama.cpp and the QVAC SDK. These are file sizes, not complete memory requirements. Leave room for the runtime, conversation state and the rest of your application.
| GGUF version | Listed size | Practical starting point |
|---|---|---|
| BF16 | 1.52 GB | Unquantized reference |
| Q8_0 | 0.81 GB | Vendor-recommended balance |
| Q4_K_M with calibration | 0.53 GB | Tighter storage and memory budgets |
QVAC reports nearly unchanged average scores for Q8_0 and a lower average for Q4_K_M across ten tests. That is not a promise of equal behavior on your tool schemas. Start with Q8_0 if it fits, then compare Q4 on the tasks you actually need. ByteForward has not run these models or measured phone performance.
For Hugging Face tooling, the separate BF16 safetensors repository includes usage instructions. BotPsy retains the Qwen3.5 architecture. Keep the supplied chat template and tool-call formatting rather than treating it as a drop-in text completion engine.
The benchmark win has a boundary
In the research team’s November 11 report, BotPsy beats Qwen3.5-2B-Think on five of six tool-use and agent tests, but trails it on Claw-Eval. Its BFCL v4 score is 44.0 versus 37.7 for that larger model. BrowseComp-Plus is only 4.5 versus 2.8. These are vendor evaluations, not independent ByteForward results.
The low browsing score matters more for a research assistant than the broad “beats a larger model” headline. A relative improvement can still leave most difficult tasks unsolved. The authors also report weakening instruction following as requirements accumulate across turns. This is a reason to test the full conversation, not only the first successful function call.
Local inference is not an offline service
BotPsy can choose tools on the device, but an internet search or remote booking service still needs a connection. Your application supplies the integrations, permission checks and execution controls. Downloading the weights does not give the model access to calendars or messages, and local inference does not make every connected service private.
Treat the first integration as a bounded experiment. Use mock accounts and reversible actions. Include a missing tool, an ambiguous request and an operation that fails once. Check whether the assistant asks a useful question, recovers, or repeats the failed call. Require confirmation before consequential actions, regardless of where inference runs.
Who should try it
BotPsy is worth evaluating when a small, well-defined tool set matters more than broad knowledge or long research. It is not an established replacement for a general-purpose agent. The developers exclude safety-critical and high-stakes automated decisions from its intended use. The QVAC FAQ also says BotPsy is not intended for applications involving money, personal information, health or other sensitive topics, despite the same page promoting calendar and messaging examples. This is the developer’s intended-use guidance, not an additional Apache 2.0 license restriction. For a contrasting deployment scale, our Nex-N2.6 guide covers a much larger agent model and why a strong tool-selection result still does not establish end-to-end reliability.
Measure completed tasks, wrong actions, elapsed time and peak memory on your target device. Keep quantization, prompts and permissions fixed between comparisons. The adoption decision should follow those results, not the smallest download or a single leaderboard win.
