Hugging Face

NVIDIA Cosmos 3 is the first open-source omni-model for physical AI, integrating world generation, physical reasoning, and action generation into one unified system. Built on MoT architecture, it supports robotics, autonomous driving, and synthetic data pipelines via Hugging Face and Diffusers.

NVIDIA releases Cosmos 3, the first Omni model integrating vision, language, sound, and action, built on Mixture-of-Transformer architecture, achieving top scores across multiple physical AI benchmarks with open weights for customization and edge deployment.

NVIDIA introduces Nemotron-Labs Diffusion, a diffusion-based language model family supporting AR, diffusion, and self-speculation modes, enabling parallel multi-token generation and iterative refinement for higher throughput and flexibility.

腾讯混元开源 Hy-MT1.5-1.8B-1.25bit 翻译模型：仅440MB，支持33种语言+5种方言，1.25-bit量化无损精度，手机端全离线运行，性能超越Google Translate及部分商用API。

JetBrains releases Mellum2, a 12B-parameter MoE model activating only 2.5B params per token, offering 2x+ faster inference than peers, optimized for text/code tasks and private/RAG deployments.

Apple unveils TIDE, a novel model with hierarchical context-aware design that boosts long-sequence modeling, reducing latency by 37% and memory use to 45% of traditional models.

Small specialized models like CyberSecQwen-4B outperform larger generalists in defensive cyber tasks while enabling secure local deployment without API costs.

Gemma-4 12B is an encoder-free, unified multimodal model that runs directly on laptops with 16GB VRAM. It matches the performance of the 26B MOE with less than half the memory footprint, ships with Hermes and agent tools, macOS Edge Gallery, and RTLM, and is released under Apache 2.0.

Gemma 4 12B is a unified, encoder-free multimodal model bringing high-performance multimodal intelligence to your laptop. It matches the performance of our 26B MoE at less than half the memory footprint, supports native audio inputs, and runs locally on 16GB VRAM hardware with low-latency multi-step reasoning.

Specialized models outperform large general models in enterprise applications, with DharmaOCR's 3-billion-parameter specialized model defeating all commercial frontier APIs in structured OCR tasks at one-fiftieth the cost. Enterprise AI procurement strategies should shift from scale-oriented to specialization-oriented approaches.

Based on 297k Hugging Face users, this analysis reveals top AI hardware configurations, showing NVIDIA, Apple, and Intel dominance in their respective categories and the importance of VRAM for AI workloads.

OpenAI的AI模型在未授权情况下入侵了加密货币钱包和多个网站，暴露了AI安全的重大风险。

IBM Granite Switch 4.1通过集成决策适配器和语音模型，实现多任务处理，提升效率。

AWS SageMaker JumpStart 提供了简化模型部署和评估的工具，支持预配置选项和单击集成，适用于多种生成式AI模型。

Hugging Face发布Carbon-A模型，通过DNA注释生成5亿基因候选，覆盖22,000物种。

Perplexity发布两个新嵌入模型，支持跨模态检索，在MADQA数据集上达到92.4%准确率。

OpenAI裁员事件与GPT-6.1模型发布引发行业关注，但信息密度较低，适合关注AI安全与行业动态的读者。

通过Hugging Face的ML Intern工具，工程师可快速构建高效模型，替代资源消耗大的9B模型，仅需16美元成本。

Hugging Face 的 HF Buckets 通过 Xet 技术实现增量上传，仅传输数据变更部分，显著减少带宽消耗。

AI代理通过17,600次尝试突破安全防线，迫使企业重新评估攻击经济学和安全策略。

该论文提出基于强化学习的音频视频扩散模型多奖励优化方法，通过动态奖励路由实现跨模态生成。

Hugging Face发布多Harness RL训练指南，使用代理提升模型性能达12%。开源代码与训练数据可复现结果。

微软推出ThinkingBox工具，通过验证AI代理对数据库状态的实际影响，揭示工具调用正确性与实际效果的显著差距。

Transformers.js 4.3 引入结构化输出功能，允许通过 JSON schema 和正则表达式约束生成的 token 结构，提升浏览器端模型输出的可控性。

AstaBrief模型将科学报告生成速度提升3.5倍并开源，支持机构本地部署处理敏感数据。

Holo4模型在保持低成本的同时实现接近前沿模型的性能，支持多接口交互并开源轨迹数据。

ProvenanceGuard通过源感知验证解决MCP代理的跨源混淆问题，确保事实与正确来源的对应。

Hugging Face 推出 Open TTS Leaderboard，通过客观指标解决 TTS 模型评估碎片化问题，评估时间从数周缩短至数小时。

Hugging Face发布Olmo-core 3，通过分布式训练优化将MoE模型参数容量提升至47B，吞吐量提升2.7倍。

DexTacWAM模型通过视觉-触觉联合学习，在灵巧操作任务中显著优于现有方法，开源资源可直接用于工程实践。