OpenAI introduces PaperBench, a benchmark measuring AI agents' ability to replicate state-of-the-art AI research. The test evaluates whether models can autonomously implement complex scientific papers.
Google DeepMind announces Gemini 2.5, its most capable AI model, now featuring built-in thinking capabilities.
OpenAI releases image generation capability for GPT-4o, significantly more capable than DALL·E 3, with photorealistic output and image-to-image transformation features.
OpenAI releases next-generation audio models in its API, enabling developers to control text-to-speech style with natural language instructions (e.g., "talk like a sympathetic customer service agent"). New customization level for voice agents.
Google DeepMind introduces Gemini Robotics and Gemini Robotics-ER, AI models designed for robots to understand, act and react in the physical world.
Google releases Gemma 3, an open multimodal and multilingual LLM with long context support. The model combines vision capabilities, multilingual processing, and extended context window, available via Hugging Face.
Hugging Face releases LeRobot, the world's largest open-source self-driving dataset. The platform includes sensor data, video feeds, and annotations to train vision and robotic control models.
Cohere introduces Aya Vision, a multimodal multilingual model processing images and text across 119 languages. The model combines vision and language understanding for image captioning, visual question answering, and document analysis tasks in low-resource languages.
Google DeepMind makes Gemini 2.0 Flash-Lite generally available in production via Gemini API, in Google AI Studio and Vertex AI for enterprise customers.
Hugging Face releases SmolVLM2, a lightweight multimodal vision model capable of processing videos and images. Optimized for mobile and edge devices, it provides an accessible alternative to large vision models.
Google releases PaliGemma 2 Mix, a family of instruction-tuned vision-language models based on Gemma 2. Three variants (3B, 10B, 28B) combine visual and textual capabilities for multimodal tasks. Available open-source on Hugging Face.
Hugging Face releases open-source DeepResearch, an autonomous research agent that conducts in-depth investigations on complex topics. The tool integrates web search, information synthesis, and multi-step reasoning to generate detailed reports without human intervention.
Hugging Face introduces π0 and π0-FAST, vision-language-action models for general robot control. These models unify visual perception, natural language understanding, and action generation, trained on diverse robotic data to execute complex tasks without task-specific fine-tuning.
OpenAI launches Deep Research, an agent using reasoning to synthesize online information and complete multi-step research tasks. Available to Pro users today, then Plus and Team.
OpenAI releases the System Card for o3-mini model, detailing safety evaluations, external red teaming, and Preparedness Framework assessments.
Hugging Face releases a tutorial to reproduce Deepseek R1's "aha moment" using reinforcement learning. Practical guide on training models with RL to generate step-by-step reasoning.
Hugging Face adds Vision Language Models (VLM) support to smolagents. Agents can now process images and text together for multimodal tasks.
OpenAI releases a System Card for Operator outlining multi-layered safety measures: mitigations against prompt engineering and jailbreaks, privacy and security protections, external red teaming, and safety evaluations.
OpenAI introduces an agent that uses computers by viewing the screen and controlling mouse/keyboard. The agent navigates websites, fills forms, and executes complex tasks without specialized APIs.
OpenAI introduces Operator, an AI agent capable of performing complex tasks on computers by interpreting visual interfaces and executing autonomous actions. The system uses computer vision to navigate and interact with web and desktop applications.
Hugging Face releases SmolVLM 256M and 500M, ultra-compact vision-language models. These variants drastically reduce model size while maintaining multimodal capabilities, targeting edge deployments and resource-constrained environments.
OpenAI announces the Stargate Project, a strategic partnership to build massive AI infrastructure. The project mobilizes significant investments to develop next-generation computing capacity and models.
Hugging Face integrates timm (PyTorch Image Models) directly into the transformers library. Users can load and use any timm model through the standard transformers API without additional dependencies.
Text Generation Inference now supports multiple backends: TensorRT-LLM (NVIDIA) and vLLM. This integration lets users select the optimal inference engine based on their performance and infrastructure requirements.
Sentence Transformers announces optimization enabling 400x faster training of static embedding models. The method drastically reduces computation time without sacrificing vector representation quality.
Hugging Face releases smolagents, a library for building simple AI agents that generate actions as code. The approach prioritizes clarity and controllability over complex agent systems.
OpenAI introduces a deliberative alignment strategy for o1 models, directly teaching safety specifications and reasoning over them. This approach leverages the models' reasoning capabilities to enhance safety.
Hugging Face releases Big Bench Audio, a benchmark for evaluating audio reasoning in AI models. The tool measures systems' ability to understand and analyze complex audio content beyond simple transcription.
Hugging Face introduces Bamba, a hybrid model combining Mamba2 and standard attention for efficient inference. The model reduces latency and memory consumption while maintaining performance on language benchmarks.
OpenAI releases o1, improves Realtime API, introduces a new fine-tuning method, and deploys additional developer tools.
Hugging Face models are now integrated into Amazon Bedrock, AWS's generative AI platform. Users can access open-source models directly through the Bedrock API without managing infrastructure.
OpenAI launches ChatGPT Pro, a new subscription tier providing priority access to frontier models and advanced capabilities. The service aims to broaden adoption of cutting-edge AI among professional and creative users.
OpenAI releases the System Card for o1 and o1-mini, detailing safety evaluations, external red teaming, and frontier risk assessments conducted before launch according to the Preparedness Framework.
Google releases PaliGemma 2, a new family of open-source vision-language models. These models combine visual and textual capabilities for multimodal tasks, with versions optimized for different use cases.
OpenAI enables vision fine-tuning for GPT-4o. Trained models better recognize map elements (roads, buildings, landmarks) with fewer errors. Use case: improved mapping and navigation services.
Hugging Face introduces Judge Arena, a benchmark to evaluate LLMs' ability to serve as evaluators. The system tests how different models judge the quality of other LLM outputs, measuring their reliability as automated judges.
OpenAI integrates web search directly into ChatGPT. Users get fast, timely answers with links to relevant sources without leaving the interface.
OpenAI introduces SimpleQA, a factuality benchmark measuring language models' ability to answer short, fact-seeking questions.
Cohere and Hugging Face introduce Aya Expanse, a multilingual language model covering 119 languages. The model improves performance on translation, text generation, and reasoning tasks across underrepresented languages, with public benchmarks released.
OpenAI simplifies and stabilizes continuous-time consistency models, achieving sample quality comparable to leading diffusion models while requiring only two sampling steps.