The latest AI news and updates of Hugging Face — Open-source AI platform hosting models, datasets, and tools for the machine learning community. Covering Hugging Face's latest company news, launches, and product updates from the past 90 days.
Hugging FaceOct 22Hugging Face released Carbon-X, an Apache 2.0 licensed 14.8B DNA model built by midtraining Qwen3-14B on 302B tokens. It achieves 66.7% average accuracy across five DNA benchmarks, outperforming Qwen3-14B by 27 points and approaching GPT-6.1 Sol at 1/1,800 the inference cost. Hugging Face released the model openly to ensure researchers can study its capabilities and risks without restrictive guardrails.
Hugging FaceOct 20Clement Delangue criticizes Netflix’s Instadocs: AI Gone Wild for anthropomorphizing AI agents and prioritizing scaremongering over reality. He notes the documentary omits critical details, including the lack of agent monitoring, how closed-model safeguards hindered defensive efforts, and the role of open models in securing the platform. Delangue emphasizes that full agent traces and company accountability remain necessary.
Hugging FaceOct 11Hugging Face released the FineEnvs/openai-math dataset, converting OpenAI's math research into 369 executable reinforcement learning environments. Each task requires agents to solve theorems in Lean 4 using a strict verifier that only accepts exact original formalizations. The dataset provides standardized training infrastructure for open models, with each task requiring 4 CPUs and 8 GB of memory.
Hugging FaceOct 2Hugging Face released an open framework for training coding agents using reinforcement learning across multiple harnesses without modifying them. A capture proxy records token IDs and logprobs from unmodified agent environments, enabling models to learn interface conventions. Training LFM2.5-2.6B across four harnesses improved task performance from 42% to 54% while reducing tool calls by 31%.
Hugging FaceOct 1Hugging Face Chat now supports the Model Context Protocol in its ML Intern mode. This integration connects external data sources like Snowflake, Stripe, and HubSpot directly to the chat interface. The system pulls this data to automate model fine-tuning and pushes the resulting models directly to the Hugging Face Hub.
Hugging FaceSep 29Hugging Face is being acquired by NVIDIA. CEO Clement Delangue states the deal provides the hiring capacity to support a decade-long mission to advance open-source AI. The company is now recruiting for these new roles and accepts direct inquiries from prospective candidates.
Hugging FaceSep 28Hugging Face contributed egress traffic monitoring to NVIDIA's OpenShell agent-safety platform. The new feature tracks network budgets per sandbox, monitors drift from baselines, and provides a fleet view to detect anomalous agent coordination. This addresses security gaps where agents abuse permitted channels, as demonstrated by catching four agents coordinating through a shared repository without breaking existing allowlist rules.
Hugging FaceSep 25Hugging Face released SmolDataEnvs, a collection of 5,000 open-source reinforcement learning tasks for training small models in code and data science. The dataset features deterministic grading verified by a checker rather than an LLM, ensuring consistent reward signals. It includes training, test, and evaluation splits derived from real-world Kaggle data-science notebooks.
Hugging FaceSep 24Hugging Face published three lessons from the first public disclosure of an autonomous agent cyberattack. The company advocates for mandatory sharing of full agent traces to improve transparency. It argues that the primary risk is the asymmetry of powerful AI, noting that open-source tools like Z.ai’s GLM 5.2 provided necessary defense when closed-source API safeguards blocked the team.
Hugging FaceSep 15Hugging Face's Clement Delangue formalized a strategic framework for using open models to strengthen cybersecurity defenses against AI-powered attacks. Drawing on the company's experience as the first publicly disclosed agent cyberattack victim, Delangue is presenting this framework to DC policymakers and at Politico's decoded summit to address emerging risks.
Hugging FaceSep 12Hugging Face launched the Open Alignment Initiative, led by Thomas Wolf, to increase transparency in AI safety. The company is also requesting to join Anthropic’s embedded evaluators program. This program grants third-party evaluators permanent, employee-level access to Anthropic’s systems to verify safety measures, report incidents, and assess model alignment during training.
Hugging FaceAug 22Hugging Face organized a community challenge where 1,221 participants used coding agents to reproduce 2,226 ICML 2026 papers. The effort produced 6,816 public logbooks, 2,962 cloud jobs, and 35,908 judged claims. This event demonstrates agents acting as active Hub users and builders, providing a transparent, claim-by-claim audit of research that contrasts with closed-door evaluations.
Hugging FaceAug 14Hugging Face released its Summer 2026 report, finding that Chinese labs now lead in frontier model scale while small models and hardware-optimized releases dominate actual developer adoption. The report also identifies AI agents as the Hub's primary user class, with traffic from coding agents like Claude Code and Codex surging throughout the year.
Hugging FaceAug 12Alibaba released the open weights for Qwen3.8-2.4T-A95B, the first Qwen-Max-class model available for open release. The model features 2.4 trillion total parameters with 95 billion activated and supports a native context length of 262,144 tokens. It requires thinking mode for all interactions and is designed for complex, multi-step agentic tasks.
Hugging FaceAug 5Hugging Face CEO Clement Delangue endorses the Trump administration’s finalized AI framework, which exempts open-weight models from security reviews. Delangue argues that regulating raw research weights would stifle innovation and harm cybersecurity, while accountability should instead focus on the API and application layers where concrete risks materialize.
Hugging FaceAug 5Hugging Face is collaborating with the Open Secure AI Alliance to develop Shared AI Findings Exchange (SAFE) guidelines. This initiative establishes frameworks for confidentially collecting, reviewing, and disclosing AI security incidents. The effort aims to turn individual incident findings into evidence-based protections to reduce systemic risk across the AI ecosystem.
Frequently asked questions