Open-source ML community and model hosting platform offering the Transformers ecosystem and compute services like HF Jobs.
llama.cpp and its underlying tensor library ggml have officially reached version 1.0.0, establishing a semantic versioning scheme that flags breaking changes to the public C API. The project now supports over 150 model architecture variants and mainstream speculative decoding methods, backed by expanded self-hosted CI runners across NVIDIA, AMD, Qualcomm, Arm, IBM and Intel hardware.
llama.cpp has officially released v1.0.0, establishing a stable baseline 3.5 years after inception. Built on quantization, optimized backends and a portable C/C++ implementation, it runs multimodal LLMs from consumer hardware to enterprise servers, serving as a key foundation for local and private AI inference.
Jensen Huang has resigned as Nvidia's president, CEO and director effective immediately, with CFO Colette Kress stepping in as interim CEO. Hugging Face CEO Clement Delangue publicly lamented the departure and vowed to defend open-source AI more firmly. The move lands amid Nvidia's roughly $12.9 billion acquisition of Hugging Face, which the DOJ has already scrutinized over competition concerns, leaving the AI compute and open-source landscape facing major uncertainty.
After Nvidia's $12.9 billion acquisition of Hugging Face, CEO Jensen Huang departed on Thursday, prompting Hugging Face CEO Clément Delangue to call it a "really sad day" while vowing to keep fighting for open-source AI. The exit follows Nvidia's second-largest acquisition ever and highlights governance and roadmap uncertainty for the open-source AI platform amid Big Tech consolidation.
Hugging Face published a comprehensive fine-tuning guide for the 3D generation model TRELLIS.2, shipping ready-to-run scripts and containers that work locally or on Hugging Face Jobs. Users can LoRA-fine-tune on their own 3D assets with just a few arguments after uploading a dataset to the Hub, while the guide also walks through the training pipeline for further customization.
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 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.