AI lab announcements, 14–20 July 2025

22 announcements from 8 labs, first-party only

OpenAI

OpenAI· 18 Jul

A $50 million fund to build with communities

OpenAI is launching an initial $50 million fund that supports nonprofit and community organizations, informed by the independent OpenAI Nonprofit Commission report.

OpenAISafety or policy· 17 Jul

ChatGPT agent System Card

describes an agentic model that combines research, browser automation, and code tools with safeguards.

OpenAICapability change· 17 Jul

Introducing ChatGPT agent

OpenAI introduces a ChatGPT agent that can think and act, using tools to complete tasks such as research, bookings, and slideshows with user guidance.

OpenAI· 17 Jul

Invideo AI uses OpenAI models to create videos 10x faster

Invideo AI uses OpenAI's GPT-4.1, gpt-image-1, and text-to-speech models to create videos.

OpenAI· 17 Jul

Statement from the OpenAI Board of Directors on the Nonprofit Commission Report

OpenAI's Board of Directors thanks the independent OpenAI Nonprofit Commission for their extensive work and engagement.

OpenAI· 17 Jul

OpenAI nonprofit jam

OpenAI Academy, in partnership with Walton Family Foundation and Emerson Collective, is hosting the Nonprofit Jam, a one-day nationwide event for over 1,000 nonprofit leaders across 10 locations.

OpenAI· 15 Jul

Intellectual freedom by design

OpenAI states that ChatGPT is designed to be useful, trustworthy, and adaptable, allowing users to customize it.

Hugging Face

Hugging Face· 18 Jul

Arc Virtual Cell Challenge: A Primer

This blog post serves as a primer for the Arc Virtual Cell Challenge, explaining the goal of training a model to predict the effect of gene silencing using CRISPR, describing the dataset, the…

Hugging FaceCapability change· 17 Jul

Five Big Improvements to Gradio MCP Servers

Hugging Face announced five improvements to Gradio MCP servers as of version 5.38.0: seamless local file support via a File Upload MCP server, real-time progress notifications, one-line…

Hugging FaceResearch paper· 17 Jul

Back to The Future: Evaluating AI Agents on Predicting Future Events

Hugging Face and collaborators propose FutureBench, a benchmark that evaluates AI agents on their ability to predict future events using real-world prediction markets and news.

Hugging FaceDeveloper tool· 17 Jul

Consilium: When Multiple LLMs Collaborate

Consilium is a multi-LLM platform that allows multiple AI models to discuss complex questions and reach consensus through structured debate, built during the Gradio Agents & MCP Hackathon.

Hugging FaceNew model· 16 Jul

Ettin Suite: SoTA Paired Encoders and Decoders

Hugging Face introduces the Ettin Suite, a set of paired encoder-only and decoder-only models (17M-1B parameters) trained on identical data, architecture, and recipes, achieving state-of-the-art…

Hugging FaceInfrastructure· 15 Jul

Migrating the Hub from Git LFS to Xet

Hugging Face is migrating the Hub's storage backend from Git LFS to Xet, a new storage system that scales with AI workloads.

Together AI

Together AIResearch paper· 17 Jul

Back to The Future: Evaluating AI Agents on Predicting Future Events

Introduces FutureBench, a benchmark for evaluating AI agents on their ability to predict future events, aiming to measure reasoning and synthesis skills while avoiding data contamination by using…

Together AICapability change· 17 Jul

Together AI Delivers Top Speeds for DeepSeek-R1-0528 Inference on NVIDIA Blackwell

Together AI announces rolling out NVIDIA Blackwell support for Together Inference, achieving top speeds for DeepSeek-R1-0528 inference with up to 334 tokens/sec.

Together AIStory of the weekAvailability change· 14 Jul

Kimi K2: Leading Open-Source Model Now Available on Together AI

Together AI announces the availability of Kimi K2, a 1 trillion parameter open-source model from Moonshot AI, on its platform with serverless deployment, pricing at $1.00 per 1M input tokens and…

Replicate

ReplicateAvailability change· 17 Jul

Bria is now on Replicate

Bria's suite of commercial-safe visual AI models is now available on Replicate.

Replicate· 15 Jul

How we optimized FLUX.1 Kontext [dev]

Replicate explains how they optimized FLUX.1 Kontext [dev] using TaylorSeer, a method that approximates intermediate diffusion steps by caching derivatives, enabling faster inference without quality…

LG AI Research

LG AI ResearchNew model· 15 Jul

Unveiling EXAONE 4.0, the next generation of hybrid AI

LG AI Research announced EXAONE 4.0, a hybrid AI model that integrates general NLP and reasoning capabilities, available in 32B and 1.2B sizes, and released for research and academic purposes via…

Mistral AI

Mistral AICapability change· 17 Jul

Le Chat dives deep.

Mistral AI announces a new capability for its Le Chat product, described as 'dives deep'.

xAI

xAIAvailability change· 14 Jul

Announcing xAI for Government

xAI announces a suite of frontier AI products available first to United States Government customers.

Z.ai

Z.aiModel update· 15 Jul

CogVideoX-3

Z.ai announced CogVideoX-3, an incremental upgrade to their video generation model with improved quality and new features, including support for start and end frame synthesis.