Releases

Hugging Face launched mutable, Xet-backed S3-like Storage Buckets for checkpoints, logs, traces, and intermediate AI data.

SpaceXAI released Grok 4.20 and a multi-agent variant through its platform.

Meta and World Resources Institute released a DINOv3-based global forest-canopy model with improved accuracy and detail.

Harvey evolved Workflow Builder into Agent Builder for reusable custom legal agents.

Harvey reported hundreds of thousands of agentic Deep Analysis reports and expanded its use in custom agent workflows.

Databricks redesigned AI Search with decoupled storage and compute to support billions of vectors at lower cost.

Scale Labs introduced repository-grounded rubrics for evaluating and reranking candidate software patches without executing tests.

Scale Labs introduced VeRO, a reproducible harness using versioned agent snapshots, controlled budgets, structured traces, and reference procedures.

Scale expanded SEAL into Scale Labs, a broader research organization focused on evaluation, agents, multimodal systems, post-training, deployments, and oversight.

OpenAI released GPT-5.4 and GPT-5.4 Pro for professional work, coding, computer use, and long-context tasks.

Harvey announced an embedded legal-intelligence agent for Microsoft 365 Copilot.

Modal launched Directory Snapshots for capturing and restoring sandbox filesystem state.

Modal added integrated changelog visibility and workflow improvements to its CLI.

CoreWeave announced an agreement to provide AI inference infrastructure for Perplexity.

Cohere joined Aston Martin Aramco as official generative-AI partner to support engineering and business innovation.

Scale launched SWE Atlas, initially releasing Codebase QnA to evaluate how coding agents investigate and reason about real software systems.

Microsoft Research releases Phi-4 Reasoning Vision, a 15-billion-parameter open-weight multimodal reasoning model.

Perplexity expanded Computer as the unifying interface for search, research, creation and actions.

AWS made centralized policy controls for agent-tool interactions generally available.

Google launches Gemini 3.1 Flash-Lite for low-cost, high-throughput multimodal workloads.

Moonshot published its distributed model-checkpointing engine.

Moonshot published optimized kernels for Kimi Delta Attention.

Moonshot released its expert-parallel communication library for large MoE training and inference.

Moonshot published Mixture-of-Block Attention code and research for efficient long-context modeling.

Baidu launched a general-purpose agent for multi-step productivity workflows.