Tether AI releases 0.8B on-device model that doubles its base's agent scores

single source· 1 articles · confidence: medium · first seen 2026-11-11 00:00 UTC

What this means for you

Downloadable now, in safetensors and quantised GGUFs, if you want an assistant that runs on the client. But the gains are the lab's own numbers with no evaluation date attached, and the lab itself says not to use it for anything high-stakes. Prototype on it; do not put it in front of a paying workflow yet.

Tether AI Research released BotPsy, a 0.8B-parameter model for tool calling and agentic workflows (where the model picks and calls functions instead of only answering in text), built on Qwen3.5-0.8B and small enough to run on a device. The lab says it beats every model up to 1.2B parameters it tested and roughly doubles its base on three agent benchmarks: BFCL v4 23.4 to 44.0, ACEBench 30.1 to 58.7, and τ-bench 23.2 to 51.5. No evaluation dates or harness details are given. Weights and quantised GGUFs are on Hugging Face; the lab says it is not for high-stakes use.

Models in this story

Key facts

  • ·BotPsy has 0.8B parameters and is built on Qwen3.5-0.8B source
  • ·The lab reports BFCL v4 rising from 23.4 for the base model to 44.0 for BotPsy source
  • ·ACEBench rises from 30.1 to 58.7, and τ-bench from 23.2 to 51.5 averaged over the retail, telecom and banking domains, according to the lab source
  • ·The lab says BotPsy outperforms every model of up to 1.2B parameters it tested on tool-use and agentic benchmarks source
  • ·Weights are published in safetensors and in GGUF quantisations (bf16, Q8, Q4) source
  • ·The lab states the model is not intended for high-stakes use source

What the sources say

  • Hugging Face Blog — Vendor model card for a 0.8B tool-calling assistant, with self-reported agent benchmark gains and no evaluation dates.

Sources

The original reporting. Follow these — they did the work.

← the wire