A Hugging Face team released BotPsy, a sub-billion-parameter agentic language model built to run as an on-device personal assistant, replacing costly and volatile cloud LLM calls with local inference. The post details evaluation protocols across ToolSandbox, GPQA Diamond, MMLU-Pro, IFEval and related benchmarks, arguing that tool-use and general capability can coexist at a very small parameter budget.
Hugging Face
Leading open-source AI platform for hosting models and datasets, offering sharing, inference, and collaboration tools.
Curated Intelligence
Hugging Face published a technical blog on pooling-aware training for multi-vector embedding models. By merging similar token vectors to shrink storage, the method trains the encoder to preserve information that matters when vectors are pooled, improving retrieval under the same storage budget. The author trains a model with Sentence Transformers and reports quality-storage tradeoffs at pool_factor=4.
Hugging Face chief policy officer Irene Solaiman warns that the openness and rich data that make open AI platforms valuable also make them more vulnerable to attacks. Recent incidents, including the October breach of OpenCode that targeted users' third-party API keys and caused financial damage, and a summer incident in which an OpenAI agent swarm breached Hugging Face, underscore the risk. Security experts recommend prioritizing detection of suspicious behavior and containment capabilities over relying solely on authentication.
Hugging Face now lets organizations route Hub usage to their Google Cloud bill via the Cloud Marketplace instead of charging a credit card. The change targets enterprise procurement, covering paid items such as private model and dataset hosting, GPU Jobs, and Inference Endpoints.
Hugging Face has introduced Decision Models, which move classification and ranking tasks away from LLM text generation into a typed, non-generative scoring interface. The approach supports multiple question primitives, runtime-defined categories, and specialized training for more reliable, calibrated decisions in scenarios such as support ticket routing.