Releases

Cohere proposed RCP-nDCG@10 as a retrieval-quality measure designed to align more closely with human relevance judgments.

Anthropic introduced a new life-sciences research group and laboratory and reported that Claude agents identified a novel array-associated reverse-transcriptase system with CRISPR-like repeats.

Modal reported support for up to one million concurrent Sandboxes.

Anthropic published case studies on misuse of Claude across cyber, influence, surveillance, fraud, biological, weapons, and model-distillation activity detected from December 2025 through August 2026.

SpaceXAI published LatchBio evaluation results reporting strong detection and refusal performance on adversarial biological tasks.

CoreWeave reported record MLPerf Training v6.0 results, including training DeepSeek-V3 in approximately two minutes on a large NVIDIA GB300 NVL72 cluster.

IBM introduced a system for composing model adapters into deployable specialized models.

Perplexity published details of the security architecture and safeguards built into Computer.

Artificial Analysis data placed CoreWeave first for Kimi K2.6 inference throughput and price-performance among evaluated providers.

Perplexity published its approach to building and evaluating accuracy in frontier AI systems.

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

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

Scale researchers released LHAW, a framework for generating underspecified long-horizon tasks and evaluating whether agents clarify ambiguity and recover performance.

xAI published its frontier-AI risk-management framework covering capability evaluation, safeguards and model development practices.

Perplexity released BrowseSafe research on detecting and mitigating threats in AI browsers.

DeepSeek introduced its open OCR and document-understanding research family based on visual-text compression.

Moonshot introduced Kimi Linear and Kimi Delta Attention as a hybrid linear-attention approach.

Scale released MCP Atlas to evaluate how well AI agents combine tools across real Model Context Protocol servers.

Scale released a harder, contamination-resistant benchmark using real and commercial repositories to measure software agents on complex multi-file engineering tasks.

xAI published the Grok 4 model card covering evaluations, safeguards, abuse potential and dual-use capabilities.

Google DeepMind introduces Genie 3, a general-purpose world model that generates interactive environments in real time.

xAI removed inappropriate Grok posts containing antisemitic tropes and praise for Hitler, tightened hate-speech controls and later explained that a system update had made the bot too susceptible to extremist user content.

Perplexity published work on accelerating Sonar inference through speculative decoding.

The companies submitted the largest NVIDIA Blackwell MLPerf Training v5.0 cluster and reported record Llama 3.1 405B training performance.

xAI disclosed that an unauthorized system-prompt change caused Grok to inject unrelated claims about South African racial politics; it published prompts and announced stronger review and monitoring controls.

Google DeepMind introduces AlphaEvolve, a Gemini-powered evolutionary coding agent for algorithm discovery and optimization.

CoreWeave submitted the first cloud-provider MLPerf Inference v5.0 results for NVIDIA GB200, reporting more than 800 tokens per second on Llama 3.1 405B.

Anthropic published interpretability research tracing internal mechanisms underlying multilingual reasoning, planning, and unfaithful explanations.

Moonshot introduced the Kimina-Prover line of open models for formal theorem proving.

Microsoft Research and Xbox introduce Muse, a generative world and action model trained on gameplay sequences.

Scale published MASK as part of its safety research program for testing whether models knowingly state beliefs inconsistent with their internal representations.

Scale released FORTRESS to evaluate frontier-model safeguards against dual-use national-security and public-safety risks.

Anthropic launched the Economic Index to measure how AI is being used across occupations and tasks.

Meta released PARTNR, a benchmark, dataset and planning model for human-robot collaboration on household tasks.

Meta researchers introduced a non-invasive deep-learning system for decoding typed sentences from EEG and MEG brain recordings.

Anthropic published and tested Constitutional Classifiers, a safeguard approach designed to resist broad classes of jailbreaks.

Scale and the Center for AI Safety published Humanity’s Last Exam, assembled from nearly 1,000 contributors across more than 500 institutions in 50 countries.

DeepSeek introduced its reinforcement-learning reasoning model family and technical report.

Microsoft Research publishes MatterGen, a generative model that designs stable inorganic materials conditioned on target properties.

Moonshot described k1 as a long-context reinforcement-learning reasoning system.

Meta's first TRIBE brain-response model formed the foundation for its award-winning Algonauts 2025 entry; Meta's later TRIBE v2 announcement identifies the original model as 2025 work.

DeepSeek introduced its 671B-parameter V3 mixture-of-experts architecture.

Anthropic and Redwood Research published evidence of alignment-faking behavior in a controlled Claude 3 Opus experiment.

Microsoft Research introduces BioEmu for efficiently sampling diverse protein conformational ensembles.

Google DeepMind introduces Genie 2, capable of generating diverse playable 3D environments from a prompt image.

Moonshot disclosed k0-math as an experimental reinforcement-learning model for mathematical reasoning.

DeepSeek introduced the Janus family for unified multimodal understanding and generation.

Meta released research on a multimodal language model that freely mixes text and speech.

Meta introduced foundation models for video generation, personalized video, editing and synchronized audio.

Scale and research collaborators introduced EnigmaEval, a multimodal reasoning benchmark built from novel puzzle-competition problems.

Scale released MultiChallenge to measure model performance on realistic multi-turn conversation problems involving memory, instruction retention, editing, and self-consistency.

Meta introduced a research pipeline for generating high-quality 3D assets with text and texture control.

Microsoft changes Recall from broad Copilot+ PC availability to a Windows Insider preview while strengthening security and privacy protections.

DeepSeek introduced its formal theorem-proving model research line.

Anthropic demonstrated feature amplification in Claude by making a discovered Golden Gate Bridge feature unusually active.

Anthropic published a large-scale mechanistic interpretability study mapping millions of concepts inside Claude 3 Sonnet.

Microsoft Research introduces Aurora, a foundation model for weather and atmospheric forecasting across multiple environmental domains.

Meta introduced an early-fusion multimodal foundation model capable of generating and reasoning over interleaved text and images.

Anthropic published research showing how long-context models can be jailbroken using many demonstrations and proposed mitigations.

Google DeepMind introduces SIMA, an instructable agent that follows natural-language directions across multiple 3D environments.

DeepSeek introduced an open vision-language model family for real-world multimodal understanding.

Moonshot released Mooncake, a KV-cache-centric architecture for disaggregated large-model serving.

Moonshot began publishing open systems research for efficient long-context model training and serving.

Google DeepMind introduces Genie, a generative interactive environment model trained from unlabelled internet videos.

Cohere Labs' Aya initiative culminated in its first open multilingual model and dataset work spanning 101 languages.

DeepSeek introduced its open mathematical reasoning research line.

Google Research introduces AMIE, a research AI system optimized for diagnostic medical dialogue.

Meta introduced Audiobox, a foundation research model for generating and editing speech, voices and sound effects.

Meta launched Purple Llama, an open trust-and-safety initiative for generative AI.

Google DeepMind reports that GNoME identified 2.2 million new crystal structures, including hundreds of thousands of stable candidates.

Microsoft Research releases Orca 2 in 7B and 13B sizes, training smaller models to use different reasoning strategies.

Meta published Emu Edit for instruction-based image editing with precise task control.

Meta published Emu Video, a factorized text-to-video generation method.

Google DeepMind publishes GraphCast, a machine-learning model for fast, accurate medium-range global weather forecasts.

IBM Research published a memory-centric neural inference chip designed for high energy efficiency.

Google DeepMind publishes AlphaMissense predictions for the effects of millions of human missense variants.

Microsoft Research introduces Phi-1, showing that curated textbook-quality data can produce strong code and reasoning performance in a compact model.

Microsoft Research introduces Kosmos-2, linking generated language to object regions in images through grounding.

Meta introduced Voicebox, a generative speech model capable of editing, sampling and cross-lingual speech generation.

Microsoft Research introduces Orca, a 13-billion-parameter model trained from rich explanation traces produced by larger models.

Anthropic published the principles used to train and guide Claude's behavior through Constitutional AI.

Microsoft Research introduces Kosmos-1, a multimodal large language model capable of perceiving visual inputs and following language instructions.

Google Research introduces MusicLM, a hierarchical text-to-music generation model for high-fidelity audio.

Google Research presents Med-PaLM, adapting PaLM to medical question answering and clinical knowledge benchmarks.

Anthropic published Constitutional AI, a method for training a harmless assistant through AI feedback guided by explicit principles.

IBM Research disclosed an AI accelerator built for efficient enterprise inference.

Google Research introduces AudioLM for generating long-term coherent, high-fidelity audio from discrete representations.

Meta presented Make-A-Video, a system for generating short videos from text prompts.

Amazon researchers published a 20-billion-parameter multilingual sequence-to-sequence teacher model.

Meta introduced Make-A-Scene, a multimodal generative method that gives creators scene-level control over images.

Google Research introduces Imagen, a text-to-image diffusion model emphasizing photorealism and language understanding.

Google Research introduces the 540-billion-parameter Pathways Language Model and reports strong few-shot performance across hundreds of tasks.

Facebook AI and university partners introduced Ego4D, a large egocentric video dataset and benchmark suite.

Baidu published its unified framework for knowledge-enhanced language understanding and generation.

Facebook AI introduced unsupervised speech recognition using unpaired speech audio and text.

DeepMind publishes MuZero, which plans using a learned model without being told the environment's underlying rules.

AlphaFold 2 demonstrates accuracy competitive with experimental structures at CASP14, marking a major protein-folding breakthrough.

Facebook AI introduced wav2vec 2.0 for self-supervised speech learning with limited labeled data.