Tether's 0.8B agent model doubles its base on tool use
Tether AI Research released BotPsy, an open 0.8B model built on Qwen3.5-0.8B for function calling on phones and laptops.
2 minOpen-Source AI ModelsFresh · 11 Nov
Tether AI Research, posting as the qvac team on Hugging Face, released BotPsy on 11 November: a model of 0.8 billion parameters trained to call tools and run simple agent tasks on a phone or laptop rather than in the cloud. It is built on Qwen3.5-0.8B, and the weights are published as safetensors and as GGUF files in bf16, Q8 and Q4 precision.
The case the team makes is about cost and control. Frontier cloud models are expensive, their prices can move, and a provider can change or withdraw access; many routine tasks do not need frontier intelligence and can be handled locally, which also keeps user data on the device. BotPsy is pitched as a personal assistant across web search, travel, commerce, customer operations, device and vehicle control, files, calendar, email and messaging.
On the team's own benchmarks, tool-use training roughly doubled the base model's agent scores. BFCL v4 rises from 23.4 to 44.0, ACEBench from 30.1 to 58.7, and τ-bench, averaged over its retail, telecom and banking sets, from 23.2 to 51.5, all against Qwen3.5-0.8B-Think. The post says BotPsy beats every model of up to 1.2 billion parameters the team tested, and leads five of six agent benchmarks even against Qwen3.5-2B-Think, a model 2.5 times its size; only on Claw-Eval does the larger model come out ahead.
The authors are plain about the limits. A model this small suits tasks of low to medium difficulty, still makes mistakes and is not meant for high-stakes use. Tool-use training preserved and slightly improved general knowledge, but knowledge still scales mainly with parameter count; instruction following weakens as instructions pile up, and performance lags in lower-resource languages.
Retold from Hugging Face. This is a summary in our own words; follow the link for the original reporting.