BotPsy for Local Workflows: Tether Releases 0.8B Model
Tether AI Research has released BotPsy, a post-trained 0.8-billion-parameter model derived from Qwen3.5-0.8B. Open weights and GGUF versions are also intended for local tool use.
BotPsy for local workflows was released by Tether AI Research on November 11. It is a post-trained model derived from Qwen3.5-0.8B, designed for function calling—that is, selecting and calling external tools—alongside agentic workflows.
The model weights are available on Hugging Face under the Apache 2.0 license. Tether also published GGUF formats used for local inference: 4-bit Q4 quantization, 8-bit Q8, and a BF16 variant.
BotPsy for local workflows and model size
BotPsy belongs to the small language model class, with 0.8 billion parameters. The individual published files are 529 MB for Q4, 812 MB for Q8, and 1.52 GB for BF16. The smallest variant therefore also supports local deployment, where users do not need to send every input to a remote cloud interface.
This is particularly important for agentic systems. Such a model can select tools and call them with specified parameters. However, its practical quality when connected to specific tools and devices has not yet been independently documented.
Designed for less demanding tasks
Tether states that the model is primarily intended for low- to medium-complexity tasks. It also warns that BotPsy should not be used for safety-critical or high-risk automated decisions.
When working with email, files, or financial tools, it will therefore be important to monitor the model’s responses to ambiguous prompts and prompt-injection attempts. Independent safety tests in such situations are not yet available.
Benchmarks are currently assessed by the author itself
According to Tether’s own tests, BotPsy outperformed the compared models with up to 1.2 billion parameters in agentic benchmarks. Tether presents these results as top-tier for this size category, but the measurements have not yet been independently verified or reproduced.
When assessing the release, it is therefore necessary to distinguish the availability of weights and formats from the claimed performance. Independent tests should show whether the results are repeatable in the BFCL, ACEBench, τ-bench, and ToolSandbox benchmarks. The practical memory usage beyond the files’ size itself and the model’s quality when used with real tools also remain open questions.
What to watch after release
For local deployment, support in tools such as QVAC SDK, llama.cpp, and other inference environments will be important. The announcement also describes planned integrations and capabilities of the QVAC ecosystem, but these are not confirmed features of the currently released BotPsy.
The model is available for download, including GGUF quantizations. For users who want to experiment with local tool calling, it primarily offers an open small model with a 529 MB 4-bit variant. Its performance claims should currently be viewed as results published by the model’s author.
Sources
- Hugging Face Blog — BotPsy – Confirms the announcement date, model origin, intended use, limitations, methodology, and benchmarks reported by the manufacturer.
- Hugging Face — qvac/BotPsy-0.8B – Confirms the availability of the weights, derivation from Qwen3.5-0.8B, and the Apache 2.0 license.
- Hugging Face — qvac/BotPsy-0.8B-GGUFs – Confirms the release of the GGUF variants and the Q4, Q8, and BF16 file sizes.
Verified and updated: November 11, 2026 06:20
