Releases

Every model, product, paper and set of open weights the record has entered, newest first, each one linked to the announcement it came from.

235 releases recorded, 45 of them with open weights.

Date Release Maker Family Kind What it was
2026
Claude Haiku 5.5 Anthropic Claude model Cheapest, fastest small model; 75% cheaper to run than Haiku 4.5.
GPT-6 with Intelligent UI OpenAI GPT product GPT-6 rolls out to all ChatGPT tiers with interactive in-answer interfaces.
Copilot “hybrid intelligence” Microsoft product Copilot gains local file context, OS-wide actions, on-device models.
Mistral Large 4 (“Le Chonk”) Mistral Mistral model 1.05T-parameter multimodal MoE public preview; open weights due late October.
Nano Banana 2.1 Google Gemini model Updated image-generation/editing model; 4K output, search grounding.
EmbeddingGemma 2 Google open-weights 740M-parameter open embedding model; one vector space for text/code/images/video/audio.
Decisions API public beta OpenAI product New endpoint returning bounded choices/probabilities/rubric scores instead of prose.
Retail AI agents trial environment Fujitsu product Four retail AI agents plus execution platform, trial; launch June 2027.
Claude for Startups expansion Anthropic Claude product Free year of Claude Team (5 seats) plus $1,000 API credits.
Beam Reflection AI model 501B total / 23B active parameters, 1M-token effective context; Apache 2.0 weights due later this month.
Kolibri Aleph Alpha open-weights English-German MoE Transformer, 78.1B total / 3.46B active parameters, 1M-token context, Apache 2.0 weights on Hugging Face.
DeepSeek Harness v0.2 DeepSeek DeepSeek product Agent harness preview adds macOS/Windows desktop apps.
Claude Code mods Anthropic Claude product TypeScript mods that change how Claude Code behaves, shipped as plugins, in the CLI and desktop app.
Guided Vision Google product Real-time voice-driven visual assistance in Gemini Live on Android.
Claude for Government (generally available) Anthropic Claude product FedRAMP High environment for federal and state agencies, generally available after a public beta since July.
Google releases Gemini 4 Argon Google model Frontier model for software engineering and cybersecurity defense; rolls out first to vetted cyber defenders via the Fairwind Program.
GPT-6.1 Sol OpenAI GPT model Coding and computer-use model launched at DevDay 2026 at one-fifth of Astra token price.
Dots OpenAI product Always-on personal agents on their own cloud computers with 4,000+ app connections; first dot free for Pro and Business Premium.
Ultrafast OpenAI product Paid speed tier promising 8x speed in Codex and 6x in the API, launched at DevDay 2026.
ChatGPT Space OpenAI product Shared workspace for ChatGPT, launched at DevDay 2026.
Pages OpenAI product Editable documents in ChatGPT, launched at DevDay 2026.
Codex in the cloud OpenAI product Cloud-hosted Codex agents, launched at DevDay 2026.
Codex Security Cloud OpenAI product Security-focused cloud environment for Codex agents, launched at DevDay 2026.
Decisions API OpenAI product Limited-preview API for Luna classification and routing, launched at DevDay 2026.
Plugin extensions OpenAI product Sidebar and interactive-panel extensions for ChatGPT plugins, launched at DevDay 2026.
Sign in with ChatGPT OpenAI product ChatGPT identity sign-in for partner apps, launching with 16 partners.
OpenAI Marketplace OpenAI product Marketplace for enterprise commitments on partner products, launched at DevDay 2026.
Bedrock Managed Agents OpenAI/AWS product OpenAI agents managed through AWS Bedrock, announced at DevDay 2026.
Private Intelligence OpenAI product Private deployment option with Private Safety Processing, announced at DevDay 2026.
Code Review in Codex OpenAI product Automated code review inside Codex, launched at DevDay 2026.
Meetings plugin OpenAI product Meetings plugin beta for ChatGPT on macOS, launched at DevDay 2026.
Codex CLI refresh OpenAI product OpenAI refreshed the Codex command-line interface at DevDay 2026.
Pro 500 OpenAI product Launched at DevDay 2026 as a new ChatGPT subscription tier. Includes 25 times the Plus plan allowance.
Claude Sonnet 5.5 Anthropic Claude model Mid-tier model at $2/$10 per MTok; Anthropic claims 30%+ faster output and up to 30% lower cost per task than Sonnet 5.
Plugin developer portal for Claude Anthropic product Portal for submitting plugins (MCP connectors and Agent Skills) to the Claude directory, with automatic validation and safety scans.
Ray-Ban Meta Audio Meta product First camera-free Ray-Ban smart glasses; open-ear audio, 12-hour battery, from $349, shipping October 13.
Ray-Ban Meta Gen 3 Meta product Third-generation smart glasses with 12MP camera and 3K video recording, from $449, on sale the same day.
Meta VR Glasses Meta product 100-gram VR glasses with tethered compute puck and 5K micro-OLED display; $1,299.99, spring 2027.
Muse Charm Meta product Palm-sized standalone Muse AI device with fingerprint wake; price and release date not announced.
Meta Adventurer Meta product Budget AI glasses entry point from $249; pre-orders open, on sale October 23.
Meta Glasses LISA Edition Meta product New $399 edition in the Meta Glasses lineup.
Claude Marketplace Anthropic Claude product One place for more than 2,000 connectors and plugins, partner software bought with committed spend, and service partners.
Claude Opus 5.5 Anthropic Claude model Efficiency-focused update for agentic coding and large migrations, priced below the Opus 5 flagship.
GPT-6 Sol OpenAI GPT model Mid tier of the GPT-6 family at half the GPT-5.6 promotional rates, aimed at coding and high-volume work.
GPT-6 Luna OpenAI GPT model Budget tier of the GPT-6 family at half the GPT-5.6 promotional rates.
MiMo-V2.6 Xiaomi open-weights Frontier-class open-weights release, continuing Xiaomi’s open-model push.
Grok 4.7 xAI Grok model Latest iteration of the frontier model, continuing a weekly-scale release pace.
Jev TypeSafe model Model built for extended independent work; developers reported strong coding results within days.
Projects redesign (beta) Anthropic product Projects rebuilt around a coordinating conversation, with parallel threads, shared memory and a project library.
One Claude, with Claude Docs and Claude Slides Anthropic product Cowork and chat merged into one Claude; Claude Docs and Claude Slides launched, and Claude Design works inside conversations.
Claude for Financial Advisors Anthropic Claude product Plugin bundling connectors and workflow skills for advisors’ research, prep and documentation; works with BlackRock, Charles Schwab, Addepar and others.
GPT-Rosalind (out of research preview) OpenAI GPT model Life sciences research model introduced in April; left research preview and became available globally to eligible organizations via trusted access.
Fugu Max Sakana AI model Orchestration model drawing on Sakana’s largest pool of open and specialized models to date; available through an OpenAI-compatible API.
Fugu Ultra v2 Sakana AI model New version of Sakana’s Fugu Ultra orchestration model, aimed at peak performance; released alongside Fugu Max through the same API.
North Small Translate Cohere open-weights Mixture-of-experts translation model (218B total, 25B active) for 50+ languages; weights for research and non-commercial use under CC BY-NC 4.0.
DeepSeek-V4.1-Flash DeepSeek DeepSeek open-weights Smallest model in DeepSeek’s new architecture family, with native multimodal support; live on the DeepSeek API, weights on Hugging Face under MIT.
Agents API OpenAI product Public beta giving developers the harness and infrastructure behind Codex for long-running agents, with context management and subagents.
Data agent in ChatGPT Work OpenAI product Turns connected company data into answers and interactive dashboards; queries keep each account’s existing permissions.
GPT-Live-1 in the API OpenAI GPT model Voice model that listens and speaks at once, first introduced in ChatGPT, handing deeper reasoning to paired models; $0.05 per minute for the voice layer.
Muse (personal AI agent) Meta product Agent that works on a person’s behalf across their apps from a dedicated secure computer with its own browser; rolling out in the US.
ChatGPT Images 2.5 OpenAI product Sharper details, faster generation and more precise editing; rolling out to ChatGPT, ChatGPT Work and Codex users on all tiers.
GPT-Image-2.5 Flare OpenAI GPT model API image model with the quality, editing and speed improvements of Images 2.5.
GPT-Image-2.5 Sunburst OpenAI GPT model API image model with extra precision for detailed creative work, at the cost of longer generation times.
K2 Horizon MBZUAI Institute of Foundation Models open-weights Fleet of six models from 0.9B to 375B parameters, released under the Apache 2.0 license.
GPT-6 Astra OpenAI GPT model Model aimed at computer use, browsing, software engineering, cybersecurity and science; initially rolled out to a limited set of organizations.
Gemini 3.8 Flash Google DeepMind Gemini model Reasoning and coding upgrade at 3.7 Flash’s speed and cost; in the Gemini API, Gemini Enterprise and for Google AI Pro and Ultra subscribers.
Gemini 3.8 Flash Cyber Google DeepMind Gemini model Gemini 3.8 variant for vulnerability detection and automated patching, available to trusted defenders through Google’s new Fairwind Program.
Muse Spark 1.3 Meta model Update with improved performance on agentic and coding tasks; available in Muse Code and the Meta Model API.
Claude Fable 5.1 Anthropic Claude model Model for coding and knowledge work; generally available on all platforms, including Amazon Web Services, Google Cloud and Microsoft Azure.
Claude Mythos 5.1 Anthropic Claude model Same model as Fable 5.1 with different safeguards; available only through trusted access programs to vetted cyberdefenders and life scientists.
Epic integration and Healthcare Public Data plugin (ChatGPT for Healthcare) OpenAI product Lets healthcare organizations, not individual accounts, bring authorized Epic patient context into ChatGPT; a plugin adds structured data from official sources.
Gemini Omni 1.1 Flash Google DeepMind Gemini model More controllable omni-modal tier for developers.
Claude in Chrome (generally available) Anthropic Claude product Generally available on every paid Claude plan; autonomous actions checked by a safety classifier.
Built-in browser in Claude Cowork Anthropic product Claude’s own browser in the Cowork side panel, separate from the user’s; rolling out to Pro, Max and Team.
Claude memory across chat and Cowork Anthropic Claude product One memory across chat, Cowork and Claude Code, kept as editable files under Topics; sensitive topics off by default.
Gemini 3.7 Flash Google DeepMind Gemini model Core reasoning improvements.
DeepSeek V4-Pro DeepSeek DeepSeek model, open weights Official release of the larger V4 tier, 1.6 trillion parameters per the April preview.
Grok 4.6 xAI Grok model Flagship model with a 500,000-token context window, at $2 and $6 per million tokens.
Muse Glimmer Meta open-weights 30-billion-parameter agentic model published under Apache 2.0.
GPT-5.6 Luna OpenAI GPT model New default model for free-tier ChatGPT.
Qwen3.8-Max Alibaba Qwen model 2.4-trillion-parameter multimodal mixture-of-experts model; weights scheduled for release the following week.
DeepSeek V4-Flash (GA) DeepSeek DeepSeek model, open weights General release; agent benchmarks reported to beat the V4-Pro-Preview.
Claude Opus 5 Anthropic Claude model Frontier model for long-running agents and technical work.
Claude voice mode on Opus and Sonnet Anthropic Claude product Voice mode moves from Haiku alone to Opus, Sonnet and Haiku, with connected tools; beta for chat users on mobile, desktop and web.
Gemini 3.6 Flash Google DeepMind Gemini model Coding, knowledge and multimodal improvements.
Gemini 3.5 Flash-Lite Google DeepMind Gemini model Low-latency tier for high-volume automation.
iOS Simulator in Claude Code Desktop Anthropic product Pane on macOS that streams an iOS simulator so Claude can tap through your app; public beta on Pro, Max and Team. Shipped the week of 20-24 July.
Qwen3.8-Max (preview) Alibaba Qwen model 2.4-trillion-parameter multimodal model previewed at WAIC Shanghai.
Grok 4.5 xAI Grok model Released through Grok Build, the Cursor editor and the xAI API.
Grok Build xAI Grok product, open weights Open-sourced under the Apache License after a data-handling controversy.
Robostral Navigate Mistral Mistral model Mistral’s first embodied-navigation model.
Claude Code in-app browser Anthropic Claude product Built-in browser in the Claude Code desktop app; Claude can open sites and work with pages. Shipped the week of 6-10 July.
Gemini Robotics 2 Google DeepMind Gemini model Whole-body control research for robots.
Lyria 3.5 Google DeepMind model Music-generation model added to Google Flow Music.
Claude Sonnet 5 Anthropic Claude model Mid-tier model with a 1-million-token context window.
Mistral OCR 4 Mistral Mistral model Optical-character-recognition and document-understanding model.
Claude Fable 5 Anthropic Claude model General release, with flagged cybersecurity and biology requests falling back to Opus 4.8.
Claude Mythos 5 Anthropic Claude model Restricted-access release through Project Glasswing.
RTX Spark Nvidia product AI-capable chip for laptops.
Gemini computer use (3.5 Flash) Google DeepMind Gemini product Computer-use capability added to Gemini 3.5 Flash.
Nano Banana 2 Lite / Gemini Omni Flash Google DeepMind Gemini model Lighter-weight image and omni-modal model tier.
Nano Banana Pro (Gemini 3 Pro Image) Google DeepMind Gemini product Reached general availability after a November 2025 preview.
Claude Opus 4.8 Anthropic Claude model Frontier model for complex reasoning and agentic coding.
Grok Build (SWE-bench update) xAI Grok product Scored 70.8% on SWE-bench Verified.
Gemini 3.5 Flash Google DeepMind Gemini model First public release in the Gemini 3.5 family.
DeepSeek-V4 (Pro / Flash) DeepSeek DeepSeek open-weights Mixture-of-experts models with 1M-token context, MIT license, weights on Hugging Face.
A13 process node TSMC product Chip manufacturing roadmap update, a shrink of A14; production slated for 2029.
Personal Computer (Mac) Perplexity AI product Extends Perplexity’s Computer agent to act on local files and apps on Mac.
Claude Opus 4.7 Anthropic Claude model Coding-focused upgrade to Opus 4.6; adds an “xhigh” effort mode and higher-resolution vision.
Qwen3.6 Alibaba Qwen open-weights 35B total / 3B active mixture-of-experts model released under Apache 2.0, weights on Hugging Face.
Muse Spark Meta model First model from Meta Superintelligence Labs; began replacing Llama in Meta’s consumer apps.
Claude Mythos Preview Anthropic Claude model Restricted preview shared with 50+ organizations via Project Glasswing to find software vulnerabilities.
Qwen3.6-Plus / Qwen3.5-Omni Alibaba Qwen model Proprietary flagship variants, access via Qwen apps and Alibaba Cloud only.
Attie Bluesky product Claude-based chatbot app for building custom social feeds.
Lyria 3 Pro Google DeepMind model Music-generation model update for longer, more structurally aware tracks.
Voxtral TTS Mistral Mistral model, open weights Open-weights multilingual text-to-speech, zero-shot voice cloning from about 3 seconds of audio.
Mistral Small 4 Mistral Mistral model, open weights 119B-parameter mixture-of-experts model (6B active), 256k context, text and image input, Apache 2.0.
Paperclip @dotta open-weights Open-source multi-agent orchestration platform under the MIT license.
Nano Banana 2 (Gemini 3.1 Flash Image) Google Gemini model Faster image generation, 512px to 4K, up to 5-character consistency, SynthID watermarking.
Gemini 3.1 Pro Google DeepMind Gemini model Tops Humanity’s Last Exam and APEX-Agents benchmarks; built for agentic, multi-step reasoning; preview.
Indus AI models and chat app Sarvam AI open-weights India-based lab’s models described as a bet on the viability of open-source AI.
Claude Sonnet 4.6 Anthropic Claude model 1M-token context in beta; 60.4% on ARC-AGI-2; record OS World and SWE-Bench scores; default for Free and Pro.
Tiny Aya model family (Aya, Aya-Global, Aya-Earth, Aya-Fire, Aya-Water) Cohere open-weights 3.35B-parameter open-weight models covering 70+ languages; trained on 64 H100 GPUs; launched at the India AI Summit.
OpenAI begins showing ads inside ChatGPT in the United States OpenAI product OpenAI began testing ads inside ChatGPT on 9 February 2026. The trial covered logged-in US adults on the Free and Go plans.
Claude Opus 4.6 Anthropic Claude model Adds “agent teams” for parallel multi-agent work; context window expanded to 1M tokens.
GPT-5.3 Codex OpenAI GPT model Agentic coding model, 25% faster than GPT-5.2; released 15 minutes after Opus 4.6.
OpenAI macOS app for agentic coding OpenAI product Desktop app for Codex-style agentic coding, preceding GPT-5.3 Codex by three days.
Gemini Robotics (Atlas/Spot deployment) Google DeepMind, Boston Dynamics Gemini product AI technology deal to power Boston Dynamics’ Atlas humanoid and Spot robots, unveiled at CES.
Claude Cowork Anthropic Claude product Agentic desktop and web app giving Claude a sandboxed shell and file access, for non-technical users.
ChatGPT age-prediction system OpenAI product Estimates whether an account belongs to a user under 18 and restricts content accordingly.
Gemini 3 as default AI Overviews model Google Gemini product Gemini 3 became the default model powering AI Overviews in Google Search globally.
Chrome “auto browse” (Gemini 3) Google product Agentic Chrome feature for handling multi-step tasks such as booking travel.
Universal Commerce Protocol (UCP) Google product, open weights Open standard proposed for agentic commerce, covering discovery through checkout.
2025
Sora 2 OpenAI product Updated video model launched with a social app. OpenAI revised its opt-out approach for rights holders within days.
Claude Sonnet 4.5 Anthropic Claude model Coding-focused update. Anthropic said it could hold a single task for about 30 hours.
GPT-5 OpenAI GPT model OpenAI’s first unified reasoning-plus-chat model. Became the default in ChatGPT.
Grok Imagine xAI Grok product Image and video generation tool released with a permissive “spicy mode.”
Grok 4 xAI Grok model Generalist and code-specialized variants with a 256,000-token context window.
o3-pro OpenAI model Higher-compute variant of o3 for ChatGPT Pro and API users.
Claude Opus 4 Anthropic Claude model Anthropic’s top coding model at release, built for long-running agent tasks.
Claude Sonnet 4 Anthropic Claude model Mid-tier upgrade to Sonnet 3.7 with better coding and instruction-following.
Veo 3 Google DeepMind model Text-and-image-to-video model announced at Google I/O 2025.
AlphaEvolve Google DeepMind paper Coding agent that found a new, faster matrix-multiplication algorithm.
o3 and o4-mini OpenAI model General-availability reasoning models succeeding o1.
Codex CLI OpenAI product, open weights Open-source terminal coding agent with Model Context Protocol support.
Llama 4 Scout and Maverick Meta Llama open-weights Multimodal open-weight models. A third model, Behemoth, remained in training.
Gemini 2.5 Pro (experimental) Google DeepMind Gemini model Reasoning and multimodal model that topped the LMArena leaderboard at release.
GPT-4.5 OpenAI GPT model OpenAI’s largest chat model at release, later positioned as a step toward GPT-5.
Claude 3.7 Sonnet Anthropic Claude model Hybrid reasoning model released alongside Claude Code.
Claude Code Anthropic Claude product Terminal coding agent. Research preview, reached general availability on 22 May 2025.
Grok 3 xAI Grok model Reasoning model with a “DeepSearch” agent mode, trained on the Colossus supercluster.
o3-mini OpenAI model Smaller reasoning model rolled out to free and paid ChatGPT tiers.
Operator OpenAI product Autonomous agent that browses the web and completes tasks such as form-filling and ordering.
DeepSeek-R1 DeepSeek DeepSeek open-weights Open-weight reasoning model matching OpenAI o1 on several benchmarks at far lower reported cost.
2024
DeepSeek-V3 DeepSeek DeepSeek open-weights 671B-parameter mixture-of-experts model; open weights published with a technical report.
o1-preview / o1-mini OpenAI model First OpenAI “reasoning” models, using an extended chain of thought before answering.
Llama 3.1 (8B, 70B, 405B) Meta Llama open-weights 405B was the largest openly downloadable frontier-scale model to date.
Claude 3.5 Sonnet Anthropic Claude model Outperformed Claude 3 Opus and GPT-4o on several public benchmarks at launch.
Apple Intelligence Apple product On-device and “Private Cloud Compute” AI across iOS 18, iPadOS 18, macOS Sequoia.
Copilot+ PCs (with Recall) Microsoft product Windows AI-PC hardware category; Recall’s local activity index was paused after privacy criticism.
GPT-4o OpenAI GPT model Native text, audio and vision in one model; free ChatGPT tier access.
AlphaFold 3 Google DeepMind, Isomorphic Labs model Predicts structures of protein complexes with DNA, RNA, ligands and ions.
Llama 3 (8B, 70B) Meta Llama open-weights Openly downloadable weights under Meta’s community license.
Claude 3 (Opus, Sonnet, Haiku) Anthropic Claude model 200K-token context window; Opus was the first model widely seen as GPT-4’s equal.
Gemini 1.5 Pro Google Gemini model 1M-token context window in limited preview; broader availability expanded through the year.
Sora OpenAI model Text-to-video model previewed publicly; a standalone Sora.com product followed in December 2024.
2023
Gemini 1.0 (Ultra/Pro/Nano) Google DeepMind Gemini model Natively multimodal; Google said Ultra was the first model past human-expert MMLU.
Meta AI assistant and AI characters Meta product An assistant on WhatsApp, Messenger and Instagram, plus 28 celebrity personas.
Mistral 7B Mistral Mistral open-weights Apache 2.0, 7 billion parameters, released as a BitTorrent link.
DALL-E 3 OpenAI model Announced 20 September; added to ChatGPT Plus and Enterprise in October.
Stable Diffusion XL 1.0 Stability AI open-weights 3.5B-parameter open text-to-image model.
Llama 2 Meta Llama open-weights 7B, 13B and 70B, free for research and for most commercial use.
Claude 2 Anthropic Claude model 100K-token context window; public beta at claude.ai in the US and UK.
GitHub Copilot X / Copilot Chat GitHub product Added a GPT-4 chat interface to Copilot in the editor and the command line.
Bard Google product Public rollout in the US and UK, built on a lightweight version of LaMDA.
Midjourney v5 (alpha) Midjourney product The image model behind the widely shared fake photograph of Pope Francis.
GPT-4 OpenAI GPT model Multimodal (text and image) model; technical report on arXiv.
Claude Anthropic Claude product Anthropic’s first public assistant, via API and chat interface.
LLaMA Meta Llama model 7B-65B research-only models; the weights leaked publicly on 3 March 2023.
Bing Chat (new Bing) Microsoft product GPT-4-based conversational search built into Bing and Edge.
2022
ChatGPT OpenAI product A free research preview of a conversational model, reached from a web page rather than through an API key.
Cicero Meta paper Combines language model with strategic reasoning for Diplomacy
Galactica Meta model Scientific-text language model; public demo withdrawn after criticism
Optimus (“Bumble C” prototype) Tesla product First working humanoid-robot prototype shown publicly
Whisper OpenAI open-weights Speech recognition and translation; code and weights under MIT license
Stable Diffusion Stability AI open-weights Text-to-image weights published under a license that permits commercial use, downloadable by anyone with a computer to run them.
AlphaFold Protein Structure Database (200M+ update) DeepMind, EMBL-EBI dataset, open weights Expanded from about 1 million to over 200 million predicted structures
Midjourney (open beta) Midjourney product Discord-based text-to-image tool opened to the public
BLOOM BigScience open-weights 176B-parameter multilingual model; full weights published
GitHub Copilot GitHub, OpenAI product Left technical preview for general availability at $10/month
Imagen Google paper Text-to-image diffusion model emphasizing photorealism; not publicly released
Gato DeepMind paper Single 1.2B-parameter transformer performing over 600 tasks across modalities
DALL-E 2 OpenAI model Text-to-image model. Announced in April, opened to a waitlist in July, and available without a waitlist from September.
PaLM Google paper 540B-parameter dense transformer trained via the Pathways system
Chinchilla (“Training Compute-Optimal Large Language Models”) DeepMind paper 70B-parameter model shows compute-optimal scaling beats larger undertrained models
2021
Megatron-Turing NLG 530B Microsoft, Nvidia model Dense transformer language model of 530 billion parameters, trained on 560 Nvidia DGX A100 servers.
On the Opportunities and Risks of Foundation Models Stanford CRFM paper, open weights Open-access report by more than 150 researchers that named the foundation-model category.
OpenAI Codex OpenAI model GPT-3 fine-tuned on public source code. Supplies the suggestions in GitHub Copilot.
AlphaFold Protein Structure Database DeepMind, EMBL-EBI dataset, open weights More than 350,000 predicted structures covering the human proteome and 20 other organisms.
GitHub Copilot (technical preview) GitHub, OpenAI product Editor extension suggesting lines and functions of code as a developer types, built on OpenAI Codex.
Wu Dao 2.0 Beijing Academy of Artificial Intelligence model Multimodal pretrained model announced at 1.75 trillion parameters, about ten times the count in GPT-3.
DALL-E OpenAI model A 12-billion-parameter version of GPT-3 that generates images from text captions.
CLIP OpenAI model, open weights Contrastive image-text model trained on 400 million pairs. Code and weights released openly.
2020
MuZero DeepMind paper Model-based reinforcement learning across Go, chess, shogi and Atari without being given the rules.
AlphaFold 2 DeepMind model Near-experimental accuracy on two-thirds of the CASP14 target proteins. Weights and code not yet public.
Apple M1 Apple product First Apple silicon Mac chip, with a 16-core Neural Engine rated at 11 trillion operations per second.
Tesla Full Self-Driving (Beta) Tesla product City-street driver-assistance software pushed to a small early-access group of owners.
Waymo One (fully driverless, public) Waymo product Robotaxi rides with no safety driver opened to the general public in Phoenix.
AlphaFold SARS-CoV-2 protein predictions DeepMind dataset, open weights Predicted structures for six under-studied coronavirus proteins, released openly as computational predictions.
OpenAI API (private beta) OpenAI product Commercial hosted access to GPT-3 for a private group of about a dozen partner companies.
GPT-3 (“Language Models are Few-Shot Learners”) OpenAI GPT paper 175-billion-parameter language model. Access sold through an API; the weights were withheld.
2019
GPT-2 (1.5B, full) OpenAI GPT model, open weights The complete model and code, released after a nine-month staged rollout.
AlphaStar (Grandmaster) DeepMind paper Nature paper reporting Grandmaster level across all three StarCraft II races under professional conditions.
RoBERTa Facebook AI paper, open weights A retrained BERT with more data and longer training, showing the original had been undertrained.
XLNet Google, Carnegie Mellon paper, open weights Permutation-based autoregressive pretraining; reported better results than BERT on 20 tasks.
OpenAI Five OpenAI model Beat OG, the reigning Dota 2 world champions, 2-0 in San Francisco.
GPT-2 (124M, staged) OpenAI GPT model, open weights The small model and its code published openly; the full 1.5-billion-parameter version withheld, citing the risk of misuse.
AlphaStar DeepMind model Beat the professional StarCraft II players MaNa and TLO 10-1 in matches played the previous December.
2018
StyleGAN Nvidia paper, open weights A style-based generator architecture for adversarial networks, used for high-fidelity synthesis of human faces.
PyTorch 1.0 Facebook AI Research product, open weights Stable release joining the eager research mode with graph-based execution for production.
Waymo One Waymo product The first commercial robotaxi service, in the Phoenix area, with a safety driver behind the wheel.
BERT Google paper, open weights Bidirectional transformer pretraining. Became the standard baseline in language research; code and weights released in November.
GPT-1 OpenAI GPT paper, open weights Generative pretraining on unlabeled text followed by task fine-tuning, the pattern later scaled through the GPT series.
Google Duplex Google product An Assistant feature that telephones businesses to book appointments. Demonstrated at I/O; limited public testing followed.
2017
AlphaZero DeepMind paper One self-play algorithm reaching superhuman play at chess, shogi and Go from the rules alone.
AlphaGo Zero DeepMind paper A version of AlphaGo trained from self-play alone, with no human game records as input. Published in Nature.
iPhone X (Face ID) Apple product The first mass-market phone to use neural-network face authentication in place of a fingerprint reader.
OpenAI Dota 2 1v1 bot OpenAI model Trained by self-play. Beat the professional player Dendi in a one-on-one exhibition at The International 2017.
Transformer (“Attention Is All You Need”) Google paper The architecture behind nearly every large language model since. Attention only, with no recurrence and no convolution.
2016
Universe OpenAI product, open weights A platform for training agents across games, websites and desktop applications.
Amazon Go (employee beta) Amazon product A cashier-less convenience store in Seattle using computer vision and sensor fusion. Open to Amazon employees only until January 2018.
Google Neural Machine Translation (GNMT) Google paper A deep LSTM translation system, put into production on Google Translate for Chinese to English.
WaveNet DeepMind paper A generative model of raw audio waveforms, used for speech synthesis and for music.
Gym (public beta) OpenAI product, open weights A shared set of benchmark environments for reinforcement-learning research, from Atari games to simulated robots.
AlphaGo (Nature paper) DeepMind paper Deep neural networks paired with tree search. The method behind the Lee Sedol match, and the first disclosure of the 5-0 win over Fan Hui.