Google

The first model of Google's Gemini 4 generation, announced September 30, 2026 by Koray Kavukcuoglu and released first to a set of trusted cyber defend…

Google DeepMind's watermarks for AI-designed proteins, published in Nature on September 30, 2026 ("Function-preserving watermarking of AI-generated pr…

A stage-aware post-training study of which verified solutions best prepare a reasoning model for the RL stage that follows SFT. The answer it proposes…

Google's text-to-speech models of the 3.8 generation, released September 23, 2026 as two API models: Gemini 3.8 Flash TTS (gemini-3.8-flash-tts), buil…

Google Cloud AI Research's answer to overfitting in automated harness evolution. Methods that let an LLM propose and keep edits to an agent's harness…

Google's native audio-to-audio models of the 3.8 generation, released 15 September 2026 as two API models: Gemini 3.8 Live (gemini-3.8-live) and Gemin…

A Google-led framework for the exploration bottleneck in recursive self-improvement (ten of its 17 authors, including the last author, are at Google;…

Google Research's model-agnostic many-agent harness for long-horizon research in mathematics and theoretical computer science, by Vahab Mirrokni, Yuan…

"A predictive map of every possible DNA letter change in the human genome." DeepMind precomputed AlphaGenome's molecular-effect predictions for all 9…

"Increasing resolution and performance of global weather models." DeepMind's third WeatherNext generation is a Functional Generative Network mesh tran…

Google's most capable cybersecurity model, a specialist tuned from Gemini 3.8 Flash for vulnerability discovery and automated patching, succeeding Gem…

Google's third Flash release in six weeks and its most intelligent workhorse model for coding and agents, built on Gemini 3.7 Flash and shipped Septem…

The third generation of Google Research's TimesFM time-series foundation model line (original decoder-only model: ICML 2024, arXiv 2310.10688; the Git…

Google DeepMind (with Duke, Columbia and Texas A&M; senior authors include Quoc V. Le and Tao Tu) extends its Gemini-based Co-Scientist multi-agen…

Google Research (with a Virginia Tech co-affiliation for senior author Tu Vu) proposes WikiSkill, a skill-evolution framework in which agent skills co…

Google Cloud AI Research applies the agent-harness idea to the environment side of agent learning. EnvHarness is a programmable wrapper layer that res…

Google's new agentic workhorse in the Gemini 3 family, built on Gemini 3.6 Flash with algorithmic improvements to its reasoning foundation. The propri…

Generation bump for DeepMind's physical-AI stack, in three tiers: a whole-body-control VLA (demoed on Apptronik Apollo 2), the ER 2 embodied-reasoning…

Google DeepMind's latent-diffusion text-to-music model generates songs up to three minutes long with synchronized audio and lyrics. Trained in JAX on…

A specialist cyber-defence model fine-tuned from Gemini 3.5 Flash and piloted through CodeMender with governments and trusted partners. Google says it…

Google's high-volume, low-latency Gemini tier, with multimodal input, 1M context, built-in computer use, and throughput up to 350 output tokens/second…

Google's July 21, 2026 "workhorse" refresh of its Flash tier (API gemini-3.6-flash). Intelligence is held flat — AA Intelligence Index v4.3: 34,…

Google's tabular foundation model: zero-shot prediction on unseen tables via in-context learning, trained "entirely on hundreds of millions of synthet…

"Recall Is the Bottleneck for Parametric Factuality" (v2 June 19) with its companion dataset WikiProfile: 2,150 Wikipedia facts, each probed under 10…

An open-weights discrete-diffusion language model built on the Gemma 4 26B-A4B MoE base (25.2B total / 3.8B active, 128 experts + 1 shared, top-8; 256…

Encoder-free unified multimodal Gemma — a ~12B dense Transformer (11.95B per the model card) sitting between the E4B edge variant and the 26B-A4…

Standalone agent-first platform announced at I/O 2026 (May 19), backed by Gemini 3.5 Flash. Comprises a desktop app, a Go-based Antigravity CLI (repla…

Launched at I/O 2026 (May 19) as the new default model in the Gemini app, AI Mode Search, and Antigravity 2.0. 1M-token context, configurable Thinking…

New Gemini model family launched at I/O 2026: image + audio + video + text in → video out. Distinct from the LLM-tier Gemini 3.5 Flash — po…

"AI Co-Mathematician: Accelerating Mathematicians with Agentic AI." Agentic-AI workbench from Google DeepMind that lets working mathematicians interle…

Position paper from Google DeepMind examining failure modes and risks of AI agents operating in real-world environments. By Matija Franklin, Nenad Tom…

4B-parameter medical multimodal model built on Gemma 3, trained for medical text and image comprehension. MedGemma 1.5 expands to high-dimensional med…

Four-variant family released under Apache 2.0 (a notable license upgrade from prior Gemma releases). Includes a 31B dense flagship, 26B-A4B MoE (4B ac…

Current flagship. 1M input, 65K output context. 2x+ reasoning boost over Gemini 3 Pro. "Medium" thinking parameter for latency/depth tradeoff. Strong…

One of the largest openly-licensed speech corpora for Sub-Saharan African languages — ~1,250 hours of transcribed natural ASR speech plus ~235 hours o…

"Olympiad-Level Formal Mathematical Reasoning with Reinforcement Learning." AlphaZero-inspired agent that learns to prove mathematical statements in L…

Gemini-powered evolutionary coding agent for scientific and algorithmic discovery. Found the first improvement to Strassen's matrix multiplication in…

First Gemini thinking model with built-in chain-of-thought reasoning. 1M token context. Led LMArena at launch. MoE architecture. Achieved SOTA on math…

First multimodal Gemma (vision + text). 128K token context (4B+). 140+ languages. Distilled training from larger Gemini models. Gemma-3-4B competitive…

Successor to BIG-Bench Hard (BBH), replacing its 23 saturated tasks with much harder variants designed to resist frontier model capabilities. Harmonic…

Demonstrates that looped Transformers — where a k-layer model is looped L times — can nearly match a kL-layer non-looped model on reasonin…

Probabilistic weather forecasting using diffusion models at 0.25° resolution. Outperformed ECMWF ENS on 97.2% of targets (99.8% at >36h lead ti…

Diffusion-based architecture predicting joint structures of proteins with DNA, RNA, small molecules, and ions. 50%+ improvement over existing methods…

Google's first open-weight models, built on Gemini research. 2B and 7B dense Transformers. Outperformed similarly sized open models on 11 of 18 text-b…

Sparse MoE Transformer with a breakthrough 1M token context window (up to 10M in research). Near-perfect recall (>99.7%) across the full context. E…

"Solving Olympiad Geometry without Human Demonstrations." Neuro-symbolic system that solves IMO geometry problems at near-gold-medal level (25 of 30 p…

ML-based weather forecasting using graph neural networks. 10-day forecasts at 0.25° resolution in under 1 minute on a single TPU v4. Outperformed ECMW…

First Gemini family: Ultra, Pro, and Nano tiers. Natively multimodal (text, image, audio, video). 32K token context. Ultra was the first model to surp…

"Scaling Deep Learning for Materials Discovery." Graph neural network that discovered 2.2 million new crystal structures, including 381,000 novel stab…

"Faster Sorting Algorithms Discovered Using Deep Reinforcement Learning." Used deep RL to discover novel sorting algorithms at the assembly instructio…

Introduced grouped-query attention as a middle ground between multi-head and multi-query attention, with a recipe for uptraining existing checkpoints…

Google I/O 2023 successor to PaLM, "significantly smaller than the largest PaLM model but us[ing] more training compute" — the tech report valid…

"Fast Inference from Transformers via Speculative Decoding." Uses a small draft model to propose tokens verified in parallel by the large model, achie…

Pathways Language Model. 540B parameter dense Transformer trained on 6,144 TPU v4 chips using the Pathways system. 780B training tokens across multipl…

"Training Compute-Optimal Large Language Models." Found that existing LLMs were significantly undertrained: for compute-optimal training, model size a…

"Competition-Level Code Generation with AlphaCode." First AI to perform at a competitive level in programming contests. Achieved top 54.3% on Codeforc…

Showed that providing chain-of-thought reasoning demonstrations in prompts dramatically improves LLM performance on arithmetic, commonsense, and symbo…

"Finetuned Language Models Are Zero-Shot Learners." Instruction-tuned a 137B parameter model on 60+ NLP tasks described via natural language instructi…

Solved protein structure prediction at atomic accuracy. Won CASP14 with a median GDT score of 92.4 (next best: 73). Predicted structures for 200M+ pro…

Simplified Mixture-of-Experts routing for Transformers. Routes each token to a single expert (vs multiple in earlier MoE work), achieving 7x pre-train…

"An Image is Worth 16x16 Words." Applied a pure Transformer directly to sequences of image patches for classification, with no convolutional layers. A…

Composable transformations of Python+NumPy programs: automatic differentiation, vectorization, and JIT compilation to GPU/TPU via XLA. 35K+ GitHub sta…

Bidirectional Encoder Representations from Transformers. Pre-trained deep bidirectional representations by jointly conditioning on both left and right…

"Neural Discrete Representation Learning." Introduced Vector Quantised Variational AutoEncoder (VQ-VAE), which learns discrete latent representations…

The paper that introduced the Transformer architecture, dispensing with recurrence and convolutions entirely in favor of self-attention mechanisms. Mu…

First AI to defeat a human professional Go player. Combined deep neural networks with Monte Carlo tree search. Defeated European champion Fan Hui 5-0…

Open-source ML framework using dataflow graph computation, supporting CPUs, GPUs, and TPUs. 189K+ GitHub stars. Dominated deep learning research and p…

Introduced knowledge distillation: compressing a large "teacher" model into a smaller "student" by training on the teacher's soft probability distribu…

"Human-Level Control through Deep Reinforcement Learning." First deep learning agent to learn control policies directly from raw pixels via RL. Surpas…

"Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift." Normalizes layer inputs during training, enabling much…