Tether's QVAC releases BotPsy, an open-weight on-device agent model

Tether AI Research, operating under Tether Data, S.A. de C.V., published BotPsy on the official Hugging Face blog on November 11, 2026. BotPsy is a 0.8 billion parameter on-device personal assistant model built on top of Qwen3.5-0.8B, released with open weights and said to outperform a model 2.5 times its size on most of the agentic benchmarks tested.

What changed

• BotPsy has 0.8 billion parameters and is built on top of Qwen3.5-0.8B
• Released under the Apache 2.0 license for research and educational use
• Available as Hugging Face safetensors and as GGUF files (bf16, Q8, Q4) for the QVAC SDK and llama.cpp
• Scores 44.0 on BFCL v4 and 58.7 on ACEBench, up from 23.4 and 30.1 for the base Qwen3.5-0.8B model
• Scores 90.8 on IFEval for instruction-following
• Outperforms Qwen3.5-2B-Think, a model 2.5 times its size, on 5 of 6 agentic benchmarks tested

Why it matters for AI builders

A sub-1 billion parameter open-weight model tuned for on-device tool-calling shows the open-weights ecosystem pushing capable agents onto hardware too small for larger models. For developers building local-first or privacy-sensitive AI assistants, that opens the door to tasks such as calendar management and web search running entirely on a device, without a round trip to a server.

Availability and scope

Tether AI Research trained BotPsy through a 4-stage pipeline: supervised fine-tuning, multi-teacher distillation, and two rounds of RLVR. The model is explicitly scoped to low and medium-difficulty tool-calling tasks and is not intended for safety-critical use.

Sources

• BotPsy: On-Device Personal Assistant with a Sub-Billion-Parameter Agentic Language Model (Hugging Face, November 11, 2026, official)

Written with AI assistance, strictly from the sources listed above. Before it went live, automated checks verified its dates, sources and every language edition.