Releases

Moonshot released a K2 checkpoint with improved coding, instruction following and long-context performance.

Google releases EmbeddingGemma, an open multilingual embedding model optimized for on-device retrieval.

Cohere released its first dedicated machine-translation model, supporting 23 languages and private deployments.

xAI released a speedy and economical reasoning model optimized for agentic coding.

Microsoft AI releases MAI-Voice-1, its first in-house expressive speech-generation model, through Copilot experiences.

Microsoft AI begins public testing of MAI-1 Preview, its first end-to-end internally trained foundation model.

xAI released the trained Grok-2 checkpoint for community use under the Grok 2 Community License.

DeepSeek released a hybrid thinking model with improved reasoning efficiency and agent capabilities.

Cohere launched a controllable hybrid reasoning model for complex enterprise agents across 23 languages.

Meta released DINOv3, a scaled self-supervised vision foundation model for high-quality visual features.

Mistral AI updated its enterprise model tier with Mistral Medium 3.1.

NVIDIA expanded its robotics stack with updated GR00T and Cosmos models and new Omniverse libraries.

OpenAI released GPT-5 across ChatGPT and the API, including mini, nano, and Pro variants.

Anthropic released Claude Opus 4.1 with improvements to agentic tasks, coding, and reasoning.

OpenAI released gpt-oss-120b and gpt-oss-20b under an open-weight license.

IBM expanded Granite with open speech-recognition models for enterprise transcription.

Google releases Deep Think, a parallel-thinking mode based on Gemini 2.5, to Gemini Ultra subscribers.

Cohere released its first commercial multimodal Command model for enterprise document, chart, OCR and scene understanding.

Mistral AI released Codestral 25.08 and expanded its enterprise coding stack across Mistral Code and developer tools.

Oracle added xAI's Grok 4 model to OCI Generative AI.

Mistral AI released Voxtral Small and Voxtral Mini, open speech-understanding models for transcription and audio reasoning.

Moonshot published the post-trained Kimi K2 Instruct checkpoint for agentic and coding tasks.

Moonshot published the base pretrained Kimi K2 checkpoint.

Moonshot introduced Kimi K2, a one-trillion-parameter mixture-of-experts model with 32B active parameters.

Mistral AI expanded the Devstral family with Devstral Small 1.1 and a larger Devstral Medium model.

xAI launched Grok 4 with native tool use and real-time search, plus the higher-compute Grok 4 Heavy multi-agent variant.

Microsoft releases Phi-4 Mini Flash Reasoning for fast reasoning on edge, mobile and real-time systems.

Hugging Face released a fully open 3B multilingual model with 128K context, switchable reasoning, tool use, and a disclosed training recipe.

Baidu open-sourced ten ERNIE 4.5 variants spanning dense and mixture-of-experts text and vision-language models.

Google releases Gemma 3n weights and tools for efficient multimodal AI on edge devices.

Google DeepMind introduces AlphaGenome to predict regulatory effects of DNA sequence variants at high resolution.

Google DeepMind introduces an on-device vision-language-action model that runs locally on robots.

Mistral AI released Mistral Small 3.2 with improved instruction following, function calling and reduced repetition.

Google introduces Gemini 2.5 Flash-Lite as a high-throughput, low-cost model in preview.

Google makes Gemini 2.5 Pro and Gemini 2.5 Flash generally available for developers and enterprises.

Moonshot released a 72B software-engineering model optimized using reinforcement learning.

Meta released V-JEPA 2 for video understanding, physical-world prediction and zero-shot robot planning.

Mistral AI released Magistral Small and Magistral Medium, its first reasoning-focused model family.

OpenAI released o3-pro for more reliable reasoning by using additional compute.

Hugging Face released SmolVLA, a 450M vision-language-action model trained for affordable, open robotic manipulation.

Moonshot published a seven-billion-parameter formal theorem-proving model.

DeepSeek upgraded R1 with stronger reasoning, reduced hallucination and JSON/function-calling support.

Mistral AI released Codestral Embed for semantic retrieval and search across software repositories.

Anthropic released Claude Opus 4 and Claude Sonnet 4 for coding, advanced reasoning, and agent workflows.

Mistral AI and All Hands AI released Devstral Small 1, an open model designed for agentic software engineering.

Google previews Gemma 3n, an efficient multimodal open model designed for phones, laptops and edge devices.

Google launches Imagen 4 with Standard, Fast and Ultra variants for higher-quality image generation.

Google launches Veo 3 with synchronized dialogue, sound effects and ambient audio generated with video.

Mistral AI released Mistral Medium 3 for enterprise workloads, combining frontier performance with lower serving cost.

Amazon introduced a generative-AI model for coordinating robot fleets and improving fulfillment-center travel efficiency.

Moonshot published the seven-billion-parameter Kimi-Audio checkpoint.

Moonshot introduced an open audio-language model family covering audio understanding and generation.

DeepSeek released 7B and 671B theorem-proving models and ProverBench.

AWS released its most capable first-generation Nova model with a one-million-token context window.

Microsoft releases Phi-4 Mini Reasoning for efficient step-by-step mathematical reasoning on constrained hardware.

Microsoft releases Phi-4 Reasoning Plus, adding reinforcement learning for stronger and longer reasoning traces.

Microsoft releases Phi-4 Reasoning, a 14-billion-parameter model trained for complex multistep math, science and coding tasks.

Baidu released an enhanced and lower-cost reasoning model.

Baidu released an enhanced ERNIE 4.5 with improved performance and lower pricing.

OpenAI released its native multimodal image-generation model gpt-image-1 through the API.

Meta released vision encoders for images and video designed to support recognition, localization and multimodal systems.

Google introduces Gemini 2.5 Flash with hybrid reasoning and controls over how much the model thinks.

IBM added the Granite 3.3 8B instruction model to watsonx.ai and open distribution channels.

IBM added the Granite 3.3 2B instruction model to watsonx.ai and open distribution channels.

OpenAI released o4-mini, a smaller reasoning model optimized for fast, cost-efficient reasoning.

OpenAI released o3, a reasoning model with strong performance across coding, mathematics, science, and visual tasks.

Cohere introduced Embed 4 for multimodal retrieval across text, images and complex enterprise documents in more than 100 languages.

OpenAI released GPT-4.1, GPT-4.1 mini, and GPT-4.1 nano in the API with improved coding and instruction following.

AWS released a unified speech-to-speech foundation model for low-latency conversational applications.

Meta released a multimodal safety classifier for Llama 4 inputs and outputs.

Meta previewed Behemoth as a teacher model for the Llama 4 family; it was not made generally available.

Meta released Llama 4 Maverick, a natively multimodal mixture-of-experts model for general AI workloads.

Meta released Llama 4 Scout, a natively multimodal mixture-of-experts model with a 10-million-token context window.

Moonshot published the reasoning-oriented Kimi-VL A3B Thinking checkpoint.

Moonshot published the instruction-tuned Kimi-VL A3B checkpoint.

Moonshot introduced Kimi-VL, an open-weight mixture-of-experts vision-language family.

Amazon released a browser-use foundation model through the Nova Act research preview.

Google introduces Gemini 2.5 Pro Experimental as a thinking model for complex reasoning and coding tasks.

OpenAI launched native image generation in GPT-4o for precise, context-aware image creation and editing.

DeepSeek upgraded V3 with stronger reasoning, coding, writing, search and function calling.

Perplexity improved Sonar, Sonar Pro and reasoning models while reducing API prices.

NVIDIA expanded Cosmos with Reason, Transfer and Predict models plus physical AI data tools.

NVIDIA released GR00T N1, an open foundation model for humanoid robot reasoning and skills.

NVIDIA released Nano, Super and Ultra open reasoning models for building agentic AI systems.

Mistral AI released Mistral Small 3.1, adding multimodal understanding and stronger multilingual capabilities.

Baidu released its first deep-thinking reasoning model.

Baidu released its first flagship native multimodal foundation model.

Cohere released the 111B-parameter Command A, optimized for agents, RAG, tool use and multilingual enterprise workloads.

Cohere made Command A weights available for research and private deployment.

Google DeepMind introduces Gemini Robotics and Gemini Robotics-ER for embodied reasoning and robot control.

Google releases Gemma 3, a multimodal and multilingual open-model family designed to run on a single accelerator.

Mistral AI launched its first OCR model and Document AI capabilities for high-quality document parsing.

Moonshot released a preview checkpoint for formal mathematical theorem proving.

OpenAI released GPT-4.5 as a research preview in ChatGPT and the API.

IBM released updated compact time-series models for long-range forecasting.

IBM updated its open safety model for detecting risks in prompts and responses.

IBM released a compact vision-language model for document and chart understanding.

IBM released the 8B Granite 3.2 model with controllable reasoning capabilities.

IBM released the 2B Granite 3.2 model with controllable reasoning capabilities.

Microsoft releases Phi-4 Multimodal, a 5.6-billion-parameter model jointly processing speech, vision and text.