News · 2026-11-11

Tether releases BotPsy, a small model for tool-calling applications

Tether AI Research has released BotPsy, a 0.8-billion-parameter model trained for tool-calling workflows, with a downloadable Q4 file of 529 MB. The model is a developer component rather than a standalone phone assistant, and the release does not establish minimum runtime memory or measured phone speed.

Key facts

• BotPsy is a text-only model based on Qwen3.5-0.8B.
• Its Q4 weights occupy 529 MB on disk; runtime GPU memory requirements are unstated.
• The publisher reports a 31% unsafe-response rate in its prompt-injection test.
• Primary source: Tether’s release article.

The model card describes a model that chooses functions and arguments. A host application supplies the actual calendar, email or other tools, executes approved calls and returns their results. Think of the model as a dispatcher writing instructions, not the person who owns the keys to every building. The surrounding application decides which keys exist and when to use them.

The downloadable GGUF files include Q4 and other variants, with instructions for local runners. The 529 MB figure describes storage, not the working memory required to load and run a model. The dossier found no source-stated minimum GPU memory, phone model, battery measurement or end-to-end latency. It also found no independent phone test. The model-file lesson explains why weight format and runtime memory differ.

Tether reports that BotPsy improved over its base in selected tool-use evaluations after multiple training stages. Its tests are vendor-run, and the underlying Qwen checkpoint already supports tool calling. The comparison therefore does not prove that a specialized small model beats all general models. The release describes its training approach but does not supply a complete training-data and code package.

Security remains an application responsibility. The publisher reports an unsafe-response rate of 31% in its prompt-injection test and recommends a separate filter for stricter safety needs. An instruction hidden in retrieved content can try to redirect a tool-using model; prompt injection explains the mechanism. Filters alone do not establish that an application’s actual permissions prevented a harmful action.

Tether’s intended-use guidance excludes sensitive and high-stakes applications. Its “on-device” positioning is meaningful for where inference can occur, but privacy depends on the host application’s network calls, tool permissions and data handling. A small download makes experiments accessible; it does not certify a safe, fast phone assistant.

The developer has to build the assistant

A typical interaction begins when an application presents the model with a request and a list of available functions. The model chooses a function and supplies arguments; the application checks that request, runs any approved operation and returns the result. The model can then decide what to say or which step to request next. Without the host code, credentials and tool implementations, downloadable weights cannot independently read an inbox or create calendar events.

The distinction affects safety as well as convenience. A model can request that a message be sent, but the application can require a user confirmation, restrict recipients or reject invalid arguments. It can also avoid giving the model access to sensitive functions at all. Whether those protections exist depends on the integration. Local inference alone does not prevent an application from sending data to other services through tools or analytics.

Tether describes several stages of adaptation, including examples of tool use, distillation from larger teacher models and reinforcement learning with rule-based feedback. Its report gives partial credit for aspects of a tool call such as the function and arguments. That training strategy can improve performance on tasks similar to the reward and evaluation setup, but does not establish reliability on every new application or language.

The publisher reports that BotPsy surpassed selected comparison models at or below 1.2 billion parameters on six reported tool-use tests, while a selected 2-billion-parameter comparison led one of them. The comparison set is limited, and results depend on prompts, settings and the surrounding harness. The report also describes shorter reasoning or fewer turns on selected tasks; it does not provide a device-by-device latency table. Fewer generated tokens are an efficiency signal, not a measured number of seconds on a phone.

The release supplies an Apache 2.0 license for its downloadable artifacts. Developers still need to examine the actual integration, resource use and intended-use restrictions, and to test against the concrete actions their application can perform.

Primary source, verified: read the paper →

Ground Truth. (2026, November 11). Tether releases BotPsy, a small model for tool-calling applications. Ground Truth. https://groundtruth.day/news/botpsy-small-tool-calling-model.html

@misc{groundtruth:botpsy-small-tool-calling-model,
  title  = {Tether releases BotPsy, a small model for tool-calling applications},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {nov},
  url    = {https://groundtruth.day/news/botpsy-small-tool-calling-model.html}
}

Topics: open-weights · agents · local-ai · cybersecurity · prompt-injection