Hugging Face releases BotPsy, a 0.8B-parameter agentic model for on-device personal assistants
The open-source model targets secure, efficient tool use and multi-turn agentic workflows with benchmarked gains over models up to 2.5× its size.
1 source · cross-referenced
- Hugging Face published BotPsy, a 0.8B-parameter agentic language model designed for on-device personal assistants and tool use.
Hugging Face introduced BotPsy, a 0.8B-parameter agentic language model optimized for on-device deployment as a personal assistant that can call tools across domains such as web search, travel, commerce, device control, files, calendar, and email.
The model is built on Qwen3.5-0.8B and achieves state-of-the-art tool-use and agentic performance among models up to 1.2B parameters, outperforming models up to 2.5× its size on five of six benchmarks, including BFCL v4, ACEBench, τ-bench, ToolSandbox, and BrowseComp-Plus.
On multi-turn agentic tasks, BotPsy scored 82.5 on τ²-bench’s telecom subset and 46.1 on BFCL v4’s multi-turn tasks, nearly doubling its base model’s performance and surpassing Qwen3.5-2B-Think on both.
BotPsy improved instruction-following reliability, with IFBench rising from 22.0 to 55.9 and IFEval from 59.7 to 90.8, while preserving general knowledge as measured by GPQA Diamond and MMLU-Pro.
In safety evaluations, BotPsy reduced unsafe responses by 21 percentage points compared to its base model under automated red-teaming and by 15 percentage points under prompt-injection attacks, the lowest among models up to 1.2B parameters.
The model uses roughly three times fewer reasoning tokens than its base on GPQA Diamond and 2.7 times fewer on MMLU-Pro, and reduced repetitive reasoning loops from 22.9% to 1.4% and runs hitting the 32k-token limit from 28.3% to 0.2%.
BotPsy will be released under the Apache 2.0 license with open weights, available in bf16, Q8, and Q4 precisions as Hugging Face safetensors and GGUF files for the QVAC SDK.
- Nov 11, 2026 · Hugging Face
Hugging Face proposes token pooling-aware training to improve compressed multi-vector retrieval
Trust79 - Nov 11, 2026 · TechCrunch — AI
Fractile raises $720M Series C to develop AI chips and compiler stack
Trust79 - Nov 11, 2026 · The Verge — AI
Google’s Tabletop AI experiment lets users build custom collaborative apps via waitlist
Trust79