Content Deep Dive

BotPsy: On-Device Personal Assistant with a Sub-Billion-Parameter Agentic Language Model

Blog post from Hugging Face

Post Details
Company
Date Published
Author
Milan Gritta, Jihye Back, JinUk, Cho, ohjoon kwon, Patrik Lambert, Mathias Buus, and Amril Nurman
Word Count
3,572
Company Posts That Month
60
Language
-
Hacker News Points
-
Post removed?
No
Summary

Tether AI Research introduces BotPsy, a 0.8-billion-parameter language model based on Qwen3.5-0.8B that is designed for on-device function calling and low-to-medium-complexity agent workflows such as search, calendar management, messaging, files, commerce, and device control. The project argues that compact self-hosted models can reduce cloud-model costs, provider dependence, and privacy risks while providing sufficient capability for many routine tasks. According to the reported evaluations, BotPsy substantially improves on its base model and leads most tested tool-use and agentic benchmarks among models with up to 1.2 billion parameters, in several cases exceeding the larger Qwen3.5-2B-Think model, while also improving instruction following, general knowledge, reasoning efficiency, and resistance to repetitive reasoning loops. It was trained through supervised fine-tuning, multi-teacher on-policy distillation from specialized 4B-parameter teacher models, and two reinforcement-learning stages using verifiable tool-calling and general-domain rewards. The model will be released under the Apache 2.0 license in multiple formats and precisions for Hugging Face, QVAC SDK, and llama.cpp deployment, although its developers caution that it can make mistakes, performs less well on open-ended research, complex multistep tasks, Hindi and Russian, and should not be used for high-stakes or sensitive automated decisions.

Trends Found in this Post
Trend Post Mentions Total Month Mentions Posts Companies MoM
LLM 7 3,355 742 191 -49%
AI Model Fine-tuning 3 613 130 56 -29%
Reinforcement learning 3 99 55 28 -39%
Use This Data

Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.