Tether Opens BotPsy-0.8B Weights for On-Device Agents

Tether AI Research on 11 November 2026 released BotPsy-0.8B, a 0.8-billion-parameter model built on Qwen3.5-0.8B for on-device function calling and agentic tasks. In a Hugging Face blog post the same day, the team, publishing as QVAC, described BotPsy as a “personal assistant” that uses tools for work such as calendar and messaging, device and vehicle control, files and documents, web search, and similar low-to-medium-difficulty jobs. “Due to its small parameter count, it is best suited to tasks of low to medium difficulty.” The citation on the post names the legal entity as Tether Data, S.A. de C.V., doing business as Tether AI Research.

The weights are now public. The qvac/BotPsy-0.8B repository holds BF16 safetensors, and qvac/BotPsy-0.8B-GGUFs holds BF16, Q8_0 and Q4_K_M GGUF files for llama.cpp and the QVAC SDK. Both repositories are licensed under Apache 2.0. The Hugging Face API records the last modification at 06:25 GMT on 11 November 2026. One highlights line on the blog still says the model “will be released”; the live repositories are the current status.

The model card is explicit about what is missing. “Training code and training data are not part of this release.” Tether has not named a date for a consumer app.

Tether’s own scores, and the limits it states

The performance case is Tether’s. Against its Qwen3.5-0.8B-Think base model, the blog and model card say BFCL v4 rises from 23.4 to 44.0, ACEBench from 30.1 to 58.7, and τ-bench from 23.2 to 51.5, the last being a mean of retail, telecom and banking subsets. Tether also says BotPsy beats Qwen3.5-2B-Think, a model 2.5 times its size, on five of six tool-use and agentic benchmarks: BFCL v4 (44.0 versus 37.7), ACEBench (58.7 versus 55.4), τ-bench (51.5 versus 46.5), ToolSandbox (64.6 versus 41.2) and BrowseComp-Plus (4.5 versus 2.8). The exception is Claw-Eval, where Qwen3.5-2B-Think leads, 48.9 to 46.5. Those figures are Tether’s own results, from its own evaluation harness, with all models run in thinking mode. TSN has no independent scores, and they should not be reported as an established state of the art.

On safety tests, also Tether’s, the blog says automated red-teaming across six harm categories produced unsafe responses 21 percent of the time, against 36 percent for the base model. Under prompt-injection attacks, “the unsafe rate drops from 46% for its base model to 31%.” Those are still substantial residual rates, and Tether’s own caveat sits next to them: “As a small model, BotPsy still makes mistakes and is not intended for high-stakes use.” The intended-use section adds that it is “not intended for safety-critical or high-stakes automated decisions, or for applications involving money, personal information, health and other sensitive topics.” English is the recommended language.

The release sits in a crowded on-device, open-weight field. Liquid AI’s LFM2.5-24B-A2B was aimed at running agents on a user’s own device, at a much larger scale. LFM3 followed with six smaller open-weight models “for any device,” and is one of the comparisons Tether itself names in the ≤1.2-billion-parameter class. Recursive’s Flint is another recent open-weight model built to run locally, in that case for coding rather than general tool use. BotPsy’s distinctive claim is a sub-billion-parameter agent checkpoint, under Apache 2.0, with GGUF files ready for llama.cpp.

Open weights are not a high-stakes assistant

The useful news is that a 0.8-billion-parameter Qwen3.5-based agent model is now downloadable, under a commercial-friendly Apache 2.0 licence, in safetensors and GGUF form. The useful limit is everything Tether qualifies or withholds. The benchmark lifts, the comparison with Qwen3.5-2B-Think, and the red-team and prompt-injection rates are the company’s own measurements. Tether says the model still makes mistakes and is not intended for high-stakes use. Training code and training data are not in the release, and there is no consumer-app date.

For researchers and developers who want a small on-device tool-calling checkpoint, the weights are there to test. They are not a finished personal assistant, and they are not a result that should be repeated as an independent league table. The facts to watch are a later official release of the training code or data, and any consumer product that actually ships this checkpoint. Until then, BotPsy-0.8B is an open on-device agent model with Tether’s own scores attached.

Sources

1. Tether AI Research, “BotPsy: On-Device Personal Assistant with a Sub-Billion-Parameter Agentic Language Model,” Hugging Face blog, 11 November 2026. https://huggingface.co/blog/qvac/botpsy
2. Hugging Face, qvac/BotPsy-0.8B model card and Apache 2.0 licence, last modified 11 November 2026, 06:25 GMT. https://huggingface.co/qvac/BotPsy-0.8B
3. Hugging Face, qvac/BotPsy-0.8B-GGUFs, last modified 11 November 2026, 06:25 GMT. https://huggingface.co/qvac/BotPsy-0.8B-GGUFs

Related stories

Liquid AI Releases LFM2.5-24B-A2B, an Open-Weight Model for Running AI Agents on Your Own Device is the nearest recent open-weight, on-device agent release against which Tether is positioning this much smaller checkpoint.

Also relevant:

• Liquid AI Releases LFM3, Six Small Open-Weight Models for Any Device
• Recursive Releases Flint, an Open-Weight Coding Model Built to Run Locally