Skip to content
Neural Daily

ModelsNews

BotPsy: Tether's 0.8B open agent model for devices

🇭🇺 Olvasd el magyarul →

Tether's QVAC team released BotPsy, a 0.8 billion parameter on-device model for tool calling. Its own tests show it beating Qwen3.5-2B, a model 2.5 times larger, on five of six agent benchmarks, and the weights are out under Apache 2.0.

· 3 min read

BotPsy announcement cover: QVAC logo and BotPsy wordmark next to teal line icons of devices on a dark background
Image: Tether AI Research · Press kit · Source

What did Tether release?

QVAC, Tether's AI research group, introduced BotPsy on November 11. It is a 0.8 billion parameter model for tool calling and agent workflows, built on Qwen3.5-0.8B. The goal is a personal assistant that runs on a phone or laptop and handles calendar, email, files, search or device settings without the data leaving the machine (QVAC blog).

The weights are already on Hugging Face under Apache 2.0, as safetensors and as bf16, Q8 and Q4 GGUF files for llama.cpp and the QVAC SDK (model, GGUF). The blog says the release is meant for research and educational use.

What do the numbers show?

According to the company's own evaluation, BotPsy roughly doubles the agentic ability of its base model. 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. On five of six tool use benchmarks it also beats Qwen3.5-2B, a model 2.5 times its size, trailing only on Claw-Eval (46.5 vs 48.9).

Tether AI Research oszlopdiagram: BotPsy-0.8B tool use és agent benchmarkjai négy összehasonlító modell mellett
Tether AI Research · Press kit · Source

Source: Tether AI Research. Tool use and agent benchmarks in reasoning mode, 0 to 100 scale.

  • IFEval rises from 59.7 to 90.8 and IFBench from 22.0 to 55.9, the best IFBench score among the models tested.
  • 86.1 on the BFCL v4 hallucination test, which checks that the model declines when no suitable tool exists.
  • 82.5 on τ²-bench telecom, versus 69.3 for Qwen3.5-2B.
  • About three times fewer reasoning tokens than the base model on correctly answered GPQA Diamond questions (3,690 vs 11,563).
  • In automated red teaming the unsafe response rate falls from 36% to 21%, and under prompt injection from 46% to 31%.

All figures come from Tether's own testing; there is no independent verification yet. Training ran in four stages: SFT, on-policy distillation from three specialist 4B teachers, then two rounds of DAPO based reinforcement learning.

Where is it weak?

Knowledge still scales with size: 38.1 on GPQA Diamond versus 59.9 for Qwen3.5-2B. Open-ended research is weak, with 4.5 on BrowseComp-Plus, and multi-step information retrieval is its weakest area. It works beyond English but drops sharply in Hindi (24.5) and Russian (57.6). The company itself says it is not meant for decisions involving money, personal data or health.

What does this mean in practice?

If you build an agent that repeats simple tool calls, a 0.8B model can handle that work locally, with no API cost and no data leaving the device. It scores around 75 on single-call scenarios but stays below 40 on multi-step ones. Realistically, it is worth trying as a router or first pass that hands harder steps to a bigger model. We covered the small open model field earlier with Liquid AI's LFM3 family.

    By Mikola Máté Patrik

    Produced with the help of AI, reviewed by an editor. How we make our articles

    Reviewed and edited by the Neural Daily desk. About

    Related articles