AI model index

The 2026-10-17 update to DeepSeek V4.1 Flash. DeepSeek reports substantial gains across reasoning, coding, agentic, and vision benchmarks, along with shorter task times and outputs; this version has its own Hugging Face checkpoint under the MIT license.

A new full-duplex real-time voice version in OpenAI’s GPT-Live family, able to listen and respond continuously and handle pauses and interruptions.

The mini voice model in the GPT-Live-1.1 release, published as a separately named full-duplex model version.

Microsoft AI’s 2.8T-parameter MoE base model, with 80B active parameters, has completed pretraining and entered post-training.

Kuaishou launched the Kling 4.0 video-generation model on 2026-10-15. It generates up to 30 seconds per pass, with up to 15 combined image, video, subject, and voice references and up to 10 keyframes.

A full-size multimodal MoE open-weight model in the Muse family, with 816B total / 42B active parameters, 1M context, and text, image, audio, and video input.

Black Forest Labs’ 5.6B open-weight image model for text-to-image, image editing, and up to 10 reference images, with native 4K generation.

xAI’s real-time speech-to-speech model, accepting text or audio and returning text and audio responses over bidirectional WebSocket.

A legal-domain model built on GLM-5.3-Flash with specialized mid-training and reinforcement-learning post-training for legal-text workflows; API beta applications are open.

A frontier agent model in Meta’s Muse family for complex coding, long-running tasks, and computer use; its model card lists text, image, video, and audio input, text output, and a 1M inference context.

The higher-capability, separately identified Muse Wonder series model for demanding coding and agent work; its model card lists text, image, video, and audio input, text output, and a 1M inference context.

An accelerated checkpoint of Qwen-Image-2.1 for text-to-image and image editing. The official model card specifies 8 denoising steps and a 7B visual-generation architecture. Weights are public under a non-commercial research license.

OpenAI’s October GPT-6 Sol version released in ChatGPT on 2026-10-07, distinct from the September version still used in Work/Codex; official docs do not map it to a fixed API ID or separately publish a complete I/O specification.

OpenAI’s October GPT-6 Luna version released in ChatGPT on 2026-10-07, distinct from the September version still used in Work/Codex; official docs do not map it to a fixed API ID or separately publish a complete I/O specification.

A lightweight model for high-volume classification, routing, extraction, and subagent tasks, with adaptive thinking, a 1M context window, and 128K output.

An open-weight on-device multimodal embedding model mapping text, code, images, video, and audio into a shared vector space.

A stable image-model update to Nano Banana 2, retaining Flash-level efficiency while improving visual quality, text rendering, multi-turn editing, and subject consistency.

An image-generation and editing API model listed by QwenCloud on October 2, 2026. Its visual generation component is described as a 7B, 32-layer single-stream DiT, with native transparency and up to 10 reference images.

Microsoft AI’s real-time streaming transcription model, which continuously receives audio and returns incremental text while a person is speaking, followed by a final transcript.

Microsoft AI’s multilingual text-to-speech model, supporting 23 languages across 26 locales while retaining the same voice across languages.

A low-latency multilingual text-to-speech model for real-time voice agents and high-throughput workloads, supporting 23 languages.

Google’s Gemini 4 frontier model for long-horizon reasoning, software engineering, and professional knowledge work, with image and video understanding; now globally available through the API.

The Sol model in the GPT-6.1 series, with API ID `gpt-6.1-sol`; it targets complex coding, computer use, and professional work with near-GPT-6-Astra capability at lower cost.

The second model in the Claude 5.5 family and a faster, more efficient upgrade to Sonnet 5 for everyday coding, agent work, and professional documents.

A structured decision model for high-frequency business decisions, supporting classification, yes/no decisions, and scoring with probabilities and confidence; it returns structured decisions rather than natural-language text.

A real-time avatar video-generation model introduced by Meta Research on 2026-09-23, generating continuous, synchronized character video conditioned on Muse Realtime Voice speech tokens and reference media.

The Sol tier of the GPT-6 line for complex coding, reasoning, and long-running agent workflows, with a 1.05M context window and 128K output limit.

The efficiency-focused GPT-6 tier for high-volume coding, reasoning, and tool tasks, also with a 1.05M context window and 128K output limit.

The first model in the Claude 5.5 family for long-running agentic coding and knowledge work, with adaptive thinking, a 1M context window, and 128K output.

The flagship creative text-to-speech model in Gemini 3.8, emphasizing studio-grade fidelity, expressive acting, regional accents, and long-form multi-turn stability.

The fast, cost-efficient text-to-speech model in Gemini 3.8 for high-throughput production and real-time voice-agent cascades.

Tencent Hy’s image-generation and editing preview model supports text-to-image, image-to-image, and multi-turn editing; Tencent Cloud TokenHub’s current guide lists output up to 4K.

The realtime interaction tier of Qwen3.8-Omni-Flash for camera and microphone scenarios, with text/audio output and multichannel audio.

SpaceXAI’s frontier model for coding, long-horizon agents, and knowledge work, with stronger self-verification, context management, and tool workflows for extended tasks.

Xiaomi’s next-generation native full-modal agent release group with Pro, Flash, and UltraSpeed serving tiers, strengthened through large-scale reinforcement learning for self-verification, coding, computer use, and long-horizon tasks.

The flagship reasoning and agent tier in MiMo-V2.6 for complex projects, long-horizon software engineering, research, and professional workflows.

The cost-efficient full-modal reasoning tier in MiMo-V2.6, balancing speed, cost, and agentic capability.

The high-speed serving variant of MiMo-V2.6-Pro for real-time interaction and latency-sensitive production workloads.

The realtime speech tier of Qwen-Audio 3.1 for low-latency full-duplex dialogue, transcription, speech output, function calling, and web search.

StepFun’s flagship preview for agentic and professional work, with a 92-block narrow-and-deep Transformer interleaving Gated Sliding Window Attention and Sparse GQA full-attention layers; 600B total parameters, 27B active, and a 1M-token context.

The native omni-modal Flash service model in the Qwen3.8 line, with up to 1M context and text, image, audio, and video inputs.

HappyOyster’s world-exploration mode, generating a real-time video stream for entering and interacting with an open world.

HappyOyster’s real-time directing mode, changing story, camera, and environment while video is being generated.

HappyOyster’s character-acting mode, keeping a generated character’s appearance, voice, and personality consistent in real-time video.

The real-time audiovisual translation model in the Qwen3.8 family, using a Hybrid-MoE Thinker–Talker design for speaker separation, synchronized bilingual output, and contextual disambiguation.

SpaceXAI’s transcription model for batch and real-time speech recognition, improving multilingual accuracy, diarization, timestamps, and noisy real-world audio.

The real-time voice and audiovisual-interaction model in Gemini 3.8; alongside audio and text, Live Avatar can generate MP4 avatar video synchronized with dialogue.

The high-reasoning real-time audio variant of Gemini 3.8 Live, supporting background reasoning during ongoing voice interactions.

The first natively multimodal GLM-5 Flash model: 320B total, 18B active parameters, hybrid sparse and linear attention, and text, image, video, and file understanding.

A specialized GPT model for life-sciences research, strengthened for medicinal chemistry, genomics, quantitative biology, experiment design, and scientific evidence handling, with tool use in governed workflows.

Kimi Code’s K2.8 preview model, with coding and agent capability close to K3, more efficient thinking than K2.7 Code, image/video input, and up to 1M context.

OpenAI’s full-duplex voice model that listens and speaks simultaneously while delegating deeper reasoning and tool use to a backend model or agent.

The original DeepSeek V4.1 Flash snapshot, released on 2026-09-10: a 552B MoE with native text-and-image input, text output, 1M context, and up to 384K output. DeepSeek’s web, app, and API now use the separately recorded 1017 version.

Xiaomi MiMo’s invitation-only next-generation release group for professional workflows including complex reasoning, multi-agent work, coding, office tasks, web design, and computer control.

OpenAI’s next-generation image line with sharper details, more precise editing, stronger multi-turn consistency, and faster generation.

The high-throughput MAI-Image-2.6 tier for low-latency, high-volume image generation with comparable quality.

OpenAI’s sixth-generation frontier model for complex reasoning, coding, computer use, research, and document work; Astra is currently available through the API, ChatGPT Work, Codex, and GPT-6 Pro access on Pro, Business, and Enterprise.

A cyber-defense variant for the limited-access Fairwind Program, paired with CodeMender to find, verify, and fix vulnerabilities.

The stable Gemini Flash upgrade for long-horizon software engineering, autonomous agents, and complex enterprise workflows, with multimodal input, 1M context, and 65K output.

The 0902 upgraded snapshot of Qwen3.8-Max, further post-trained for coding, collaborative agents, long-horizon autonomous development, and vision while retaining a 1M context and full tool ecosystem.

Muse Spark 1.3 is built for longer-horizon coding and agent work, improving multitasking, plan correction, and tool-use efficiency in real engineering workflows.

The successor to Fable 5 for long-horizon agents, coding, research, and complex knowledge work, released with more precise cybersecurity and biology safeguards.

A real-time audio perception model in the Muse family for streaming speech recognition, diarization, endpointing, and multilingual mixed speech.

Tencent Hy’s next-generation 770B MoE flagship preview with 49B active parameters and a 1M context for coding, office analysis, game development, and research.

The GA Gemini-native model for video generation and conversational editing, accepting text, image, audio, and video with output up to 4K.

A Qwen4 architecture preview: 125B main model plus 51B N-gram embeddings, 6B active per token, extendable to 1M context.

An experimental V4 Flash API model adding image input, now superseded by DeepSeek V4.1 Flash.

Kept the GLM-5.2 base and improved complex coding, long-horizon tasks, and cyber capability entirely through scaled post-training; FP8 and BF16 weights are now public.

The 27B dense open Qwen3.8 model for local deployment, vision, coding, and tool use.

The 3.7 generation of Gemini Flash for software engineering, web development, multi-step agents, and design fidelity.

The V4-Pro production snapshot; its planned orderly sunset was superseded by an official pricing-page update that keeps the API available.

Built on Grok 4.5 with stronger long-running agents, complex-codebase work, and interactive visual projects.

The production successor in GitHub Copilot, improving code quality and token efficiency.

A purpose-trained cybersecurity model for authorized vulnerability research, exploit validation, and security testing, improving selected defensive workflows.

A 30B open model for always-on local agents, designed to run on one consumer GPU or a Mac or PC.

Microsoft’s image-generation and editing model with improved text rendering and commercial design, available in Playground and Foundry private preview.

SpaceXAI’s image generation and editing flagship, improving prompt adherence, typography, detail fidelity, and multi-image editing.

A multimodal video model, generating up to 30 seconds with as many as 20 multimodal references.

The coding-focused predecessor to Muse Spark 1.3, released alongside the multi-agent Muse Code tool.

The hosted Qwen3.8-Max API model with multimodality, tool use, and thinking mode; the open checkpoint is recorded separately.

Qwen3.8’s open checkpoint with 2.4T total and 95B active parameters, native 262,144 context, extensible to about 1M.

The V4 Flash 0731 production API snapshot, now superseded by DeepSeek V4.1 Flash.

Generates up to 30 seconds with multi-round extension, 30 image/10 video/10 audio references, and timestamp editing.

An open omni-modal generation system producing up to 15-second, 2K, 24fps video with native stereo audio.

The new Opus flagship with a default 1M context and adaptive thinking for long-running agents, research, and enterprise work.

The current default, focused on aesthetics, image quality, personalization, and fewer low-quality results.

An early-access video tier generating clips up to 20 seconds with audio, image/video references, keyframes, and continuation.

A low-latency 3.5 tier designed for high-volume automation and subagents.

Improved token efficiency, coding, and agent planning over 3.5 Flash while reducing output verbosity.

Targets directly usable visual content with 4.5K-token prompts, 10px text, multilingual rendering, and complex layouts.

A 2.8T open MoE using KDA and Attention Residuals with native vision and a 1M context window.

A fast flagship for coding, agents, and knowledge work, available in Grok Build and the API.

Improved tools, computer use, coding, and multimodal understanding with a 1M-context Meta Model API preview.

The professional-design flagship for information visualization, pixel-level interactive editing, realism, and multilingual generation.

An agentic image model that uses search and code, self-refines, and supports single- and multi-image references.

A preview Muse video model emphasizing fidelity, temporal consistency, and native audio; not yet generally available.

The final Hy3 used product feedback to improve agents, coding, and output reliability.

A new Sonnet generation with a default 1M context, adaptive thinking, and stronger agent capability closer to Opus.

The fastest and lowest-cost high-throughput model in the Nano Banana family.

A three-tier series—Sol, Terra, and Luna—covering flagship, balanced, and cost-efficient capability levels.

The current Seed line, improving long-running general agents, end-to-end coding, and scientific workflows.

Extended context to 1M and optimized for large codebases, long documents, and project-scale agent work.

Improves image-to-video motion, physics, and audio, with general availability in the Imagine API.

Kimi’s open model for coding and agent tasks, improving long-horizon coding, instruction following, and reasoning efficiency.

An experimental Gemma 4 text-diffusion model exploring lower latency through parallel block generation.

A Mythos-class model made generally available with conservative safeguards for software engineering, research, and complex knowledge work.

The current Ray model for fuller creative control and production API workflows.

A unified encoder-free multimodal model for laptops, filling the gap between E4B and 26B.

Microsoft’s efficient coding model trained from scratch and introduced in GitHub Copilot, not distilled from another model.

Microsoft’s model for complex reasoning, available in public preview through Microsoft Foundry.

MiniMax’s current open flagship with 1M context, native image/video understanding, coding, and computer use.

A high-efficiency vision-language Flash model for real-world agents, with about 198B total parameters, about 11B active, 256K context, image understanding, and tool orchestration.

Improved agent judgment, tool efficiency, computer use, and long-session collaboration over Opus 4.7.

MAI’s unified image generation and editing model, improving text rendering, localized edits, and identity consistency.

The first public model in Google’s 3.5 generation, combining frontier intelligence with low-latency action in the Flash tier.

Trained from an ERNIE 5.0 submodel structure, reducing model size and training cost while improving reasoning, agents, and creative writing.

The low-latency, high-throughput, cost-efficient tier in the Gemini 3 family.

The everyday fast tier of GPT-5.5, improving accuracy, concision, and use of personalized context.

The original Gemini-native video generation and conversational editing model, with a model card first published in May 2026 and an API preview in June.

The audio-reasoning research update to Step-Audio-R1, with a technical report and accompanying evaluation benchmarks released on 2026-04-29.

StepFun’s real-time image generation and editing service, recorded as live on 2026-04-29 in the official repository.

The high-capability DeepSeek V4 preview: a 1.6T-total, 49B-active MoE with a 1M context window.

The efficient V4 preview with 284B total, 13B active parameters, and a 1M context window.

The first Hy3 preview after Tencent rebuilt its training stack, combining fast/slow reasoning with sparse MoE.

A 1M-context frontier model for complex knowledge work, computer use, coding, and scientific research.

MiMo’s current general model combining native full-modal understanding, 1M context, and agent execution.

The high-intensity MiMo-V2.5 agent tier for long-range reasoning, coding, and complex-task efficiency.

The 27B dense Qwen3.6 tier for local deployment and general multimodal agents.

The current flagship for 2K production assets, complex layouts, multilingual text, and high-fidelity editing.

The open successor to K2.5 with stronger long-horizon coding, general agents, vision, and Agent Swarm.

Focused on hard software engineering, long-run consistency, higher-resolution vision, and self-verification.

The efficient Qwen3.6 MoE tier, using 3B active parameters for multimodal, coding, and tool tasks.

A GPT-5.4 cybersecurity model for vetted defenders, with a lower refusal boundary for legitimate vulnerability research and binary reverse engineering; scheduled for retirement under the official notice.

Added native 2K HD, roughly 4–5× faster standard generation, and stronger text and detail retention.

The first Muse model, with native multimodal reasoning, tools, visual chain of thought, and multi-agent orchestration.

A GLM-5 update for long-horizon work, designed to execute autonomously for hours within one task.

Unified audio text-to-video, first/last frames, continuation, references, and editing in a hosted 1080p line.

The high-quality dense Gemma 4 flagship for local reasoning, coding, and agent workflows.

The Gemma 4 MoE activating only 3.8B parameters per step for local speed and intelligence density.

Wan’s image model for image sequences, multi-reference, interactive editing, and output up to 4K.

The second MAI image generation, improving realism and text rendering, initially offered in Playground and selected customer APIs.

Strengthened agent teams, self-evolution, and end-to-end software productivity.

A trillion-parameter, 42B-active agent flagship with a 1M context window.

Unified text, vision, and speech perception with tools and GUI control in one agent foundation.

A fast small model for coding and subagents with text, image, tools, and a 400K context window.

The smallest GPT-5.4 tier for high-throughput extraction, ranking, and simple coding subtasks.

Unified GPT-5.3-Codex coding advances with general reasoning, tools, and professional document work.

A GPT-5.3 update for everyday conversation, focused on fewer unnecessary refusals, better search accuracy, and smoother writing.

Brings 0.5K–4K generation, thinking, and image-search grounding at Flash speed and price.

The successor preview to Gemini 3 Pro, with a separate endpoint tuned to prioritize custom tools.

A full Sonnet 4.5 upgrade across coding, computer use, long context, agent planning, and design.

The first Qwen3.5 flagship, extending the Qwen3-Next hybrid architecture into a native multimodal agent model.

Systematically improved multimodal understanding, foundation reasoning, and production agent performance.

Brought deeper reasoning and live search into lightweight image generation and editing.

Scaled to 744B total parameters with DSA, repositioning the model from code generation toward complex systems engineering.

A small, ultra-fast Codex research preview designed for real-time interactive coding.

A high-efficiency update for coding, tools, and workplace productivity.

Unified text, image, audio, and video references and editing with 15-second multi-shot stereo output.

Unified generation and editing with 1K-token complex prompts, native 2K, and more realistic detail.

A Codex model extending agentic coding into research, tool use, and end-to-end computer work.

Brought a 1M context beta to Opus with stronger long-running agents, code review, and adaptive thinking.

A release group covering Video, Omni, and Image tiers; inputs, outputs, and limits depend on the individual service.

A sparse MoE model for fast reasoning, coding, and agent workflows, with 196B total parameters, 11B active per token, and 256K context.

Grok Imagine’s native video-and-audio API for generation and editing.

A natively multimodal continuation of K2 with visual coding and an Agent Swarm of up to 100 subagents.

Brought Ray3 to native 1080p while substantially reducing latency and cost.

The instruction version of HunyuanImage 3.0, adding reasoning, image editing, and multi-image fusion, separate from the original text-to-image model.

A sub-second generation and editing model for consumer GPUs.

A compact multimodal open model in the Step 3 line at about 10B scale, focused on visual perception, complex reasoning, and human-centric alignment.

The open-weight update to Step-Audio-R1, with inference code and weights released on 2026-01-14 in the official repository.

The post-training update to NextStep-1, released on 2025-12-24 with both pretraining and post-training weights according to the official repository.

Improved real-world coding, tool use, and agent reliability.

The open successor to GLM-4.6 with continued gains in coding, reasoning, and agent tasks.

A GPT-5.2 variant optimized for long-horizon software engineering, large code changes, Windows, and context compaction.

A general agent model for search, code, GUI interaction, and complex workflows.

The low-latency Gemini 3 preview with stronger visual-spatial reasoning and agentic coding.

Used hybrid sliding-window attention and MTP for faster long-context coding and agents.

Improved precise editing, detail preservation, typography, and generation speed by roughly four times.

Added joint audio-video generation, multilingual dialogue, camera control, and fuller narrative expression.

A GPT-5 series upgrade for professional work and long-running agents, offered in Instant, Thinking, and Pro tiers.

Generates videos up to 10 seconds from text or images, with dialogue, music, and ambient audio.

A general model unifying non-thinking and thinking modes with sparse attention and everyday agent capability.

A high-reasoning V3.2 variant for math, competitive programming, and longer deliberation.

Runway’s current flagship for cinematic fidelity, complex sequenced prompts, and professional HDR output.

A Kling video model combining references and editing, with text, image, and video prompts and shot extension.

An audio-native reasoning model using Modality-Grounded Reasoning Distillation so test-time compute scaling operates on acoustic information.

The second-generation 32B open-weight developer model with multi-reference editing, precise color control, and stronger realism; Max, Pro, and Flex are separate service variants rather than aliases.

Introduced effort control with stronger coding, agents, computer use, and multi-agent coordination.

A professional Gemini 3 Pro image model with reasoning, real-world knowledge, multilingual text, and 4K control.

A lighter 8.3B video model supporting text- and image-to-video with more accessible local deployment.

A 2M-context fast agent model for real-world tool use and deep research, offered in reasoning and non-reasoning variants.

The first Gemini 3 Pro preview for complex reasoning, coding, tools, and multimodal work.

Used large-scale reinforcement learning to improve style, intent understanding, collaboration, and factuality in everyday conversation.

An adaptive-reasoning API update that reduced latency on easy tasks and added apply_patch and shell tools.

Baidu’s unified multimodal foundation model jointly modeling text, images, audio, and video, first announced in November 2025.

A K2-based thinking agent trained to reason while using tools natively.

Improved body motion, micro-expressions, stylization, and motion prompts with a lower-cost Fast variant.

Unifies image generation and prompt-based editing at native 4MP resolution, initially offered in public beta.

Shifted MiniMax’s main line toward efficient coding and agent execution.

A low-latency Haiku model offering near-Sonnet-4 coding and computer-use capability at lower cost.

Improved narrative control, reference consistency, and editing, adding 1080p, 4K, and vertical output.

Microsoft’s first image model developed entirely in-house, initially shown on LMArena before entering Bing Image Creator.

Improved physics, control, and multi-shot state persistence with native dialogue and sound; web and app ended on 2026-04-26, and the API shut down on 2026-09-24.

The open model between GLM-4.5 and GLM-4.7, expanding context to 200K and improving coding and agent capabilities.

Strengthened long-running coding, computer use, and complex agents alongside the Claude Agent SDK.

An experimental model built on V3.1-Terminus, introducing DeepSeek Sparse Attention for long-context efficiency.

A native multimodal MoE image model with 80B total and 13B active parameters.

A trillion-plus-parameter Qwen3 flagship for coding, agents, and million-token professional work.

A refined V3.1 release improving language consistency, tool use, and agent performance.

An efficient Grok 4 variant unifying reasoning and non-reasoning modes with 2M context and live search.

Added visual reasoning, self-evaluation, native HDR EXR, and draft-to-4K workflows.

An 80B-total, 3B-active architecture preview combining hybrid linear attention and ultra-sparse MoE.

Unified text-to-image and editing, using a vision-language model for world knowledge and supporting up to 4K.

Microsoft’s first foundation model trained end-to-end, introduced for public testing on LMArena as an early MAI milestone.

Applied Gemini 2.5 Flash knowledge and conversation to fast image generation and editing.

A hybrid reasoning model unifying thinking and non-thinking modes and extending API context to 128K.

An image-editing extension of 20B Qwen-Image for semantic, appearance, and bilingual text changes.

A 14B autoregressive image generation and editing model using continuous image tokens, exploring image modeling without discrete VQ tokens.

GPT-5 unified fast responses, deeper reasoning, and automatic routing, with flagship, mini, and nano variants in the API.

The lower-cost GPT-5 tier for workloads balancing reasoning quality, throughput, and price.

The smallest GPT-5 tier for classification, extraction, ranking, and lightweight subtasks.

An open reasoning model with 117B total and 5.1B active parameters per token, for customization and tool workflows.

The smaller open reasoning tier, with 21B total and 3.6B active parameters per token, targeting local deployment and lower latency.

A product upgrade to Opus 4 for coding and agent tasks, released through Claude, the API, and cloud partners.

A 20B MMDiT image foundation model focused on complex Chinese/English text rendering.

StepFun’s multimodal reasoning foundation model, co-designed around MoE and MFA/AFD system components with 321B total and 38B active parameters.

A 355B-total, 32B-active hybrid-reasoning MoE unifying reasoning, coding, and agent capabilities.

Introduced dual-expert MoE video diffusion with open 720p text, image, speech, and animation variants.

An end-to-end audio-understanding and speech-conversation release group spanning audio semantics, paralinguistic information, tool use, and multimodal retrieval.

An agentic coding MoE with 480B total and 35B active parameters; native 256K context can be extrapolated to 1M.

A 1T-total, 32B-active open MoE focused on agentic tool use and coding.

Scaled reinforcement learning beyond Grok 3 Reasoning with native tools, web search, and X search.

The open ERNIE 4.5 text MoE, with 300B total and 47B active parameters under Apache 2.0.

The open ERNIE 4.5 vision-language MoE, with 424B total and 47B active parameters, understanding images and video and producing text.

Used MatFormer and Per-Layer Embeddings to bring multimodal understanding to mobile devices.

Introduced adaptive reasoning on a 230B MoE foundation unifying text and vision.

Animates one image into five seconds, extendable to 21 seconds with end frames, loops, and motion modes.

Advanced Hailuo with native 1080p, complex motion, and stronger physics.

An open reasoning model with 1M context using hybrid linear and softmax attention.

Established Seedance with native multi-shot storytelling, 1080p, and stable complex motion.

A 12B open-weight contextual editing model for multi-turn changes, character consistency, and local control.

The Kling 2.1 series launched in late May 2025 with Standard, High Quality, and Master tiers; no shared parameter count is assumed.

The Claude 4 flagship, focused on long-running coding, complex agents, and sustained multi-hour execution.

The balanced Claude 4 tier with stronger coding, reasoning, instruction following, and parallel tool use.

The base Imagen 4 service model with up to 2K output, multiple aspect ratios, and stronger detail, spelling, and typography; Ultra and Fast are not treated as aliases.

Added native dialogue, ambience, and sound effects to Veo.

A thinking-model preview on the Kimi Open Platform: a distinct API model, not an alias of the K1.5 research model or K2 Thinking.

Xiaomi’s first open reasoning model, exploring the full pretraining-to-RL path at 7B scale.

The flagship Qwen3 MoE, combining thinking and non-thinking modes in one open model.

StepFun’s open image-editing model for object addition/removal, background changes, action changes, and style transfer through natural-language instructions.

The fourth Firefly image generation, improving realism and detail for creative workflows.

A higher-detail tier released alongside Image 4 for complex scenes and intricate imagery.

Brought GPT-4o-native image generation to the API with world knowledge, style adherence, and typography.

A hybrid reasoning model balancing speed, cost, and controllable thinking, previewed in April and generally available in June.

A reasoning model combining visual reasoning and tool use, released in ChatGPT and the API.

An efficiency-focused small reasoning model with vision and tool use.

The second Kling video generation, improving motion, semantic response, and multimodal creative control.

Improved coding, instruction following, and long-context processing, with up to one million tokens in the API.

The smaller multimodal GPT-4.1 tier balancing speed, cost, and long context.

The lowest-latency GPT-4.1 tier for high-volume tasks such as classification and completion.

StepFun’s lightweight visual reasoning model with image-text input and text output, introduced on 2025-04-08 in the official page.

A natively multimodal MoE with 17B active parameters, 16 experts, text-and-image input, and up to 10M context.

The higher-capability Llama 4 open model with 400B total parameters and 128 experts.

Added default personalization, Omni Reference, 10× draft mode, and conversational creation.

A native-2K text-to-image model improving small text, layouts, and portraits; official history dates its service launch to April 2025.

Improved subject, location, and style consistency across scenes with more direct camera control.

Uses a Thinker-Talker architecture to understand text, images, audio, and video and stream text and speech.

Google built “thinking” directly into the general Gemini generation, with emphasis on complex reasoning and coding.

An image-to-video and motion-control model built on Step-Video-T2V, with open inference code and weights.

The hosted multimodal ERNIE 4.5 product model, recorded separately from the concrete weight tiers opened in June without borrowing their parameter counts.

ERNIE’s deep-thinking branch, launched alongside ERNIE 4.5 with a focus on reasoning and tool use.

Added image input, a 128K context window, and support for more than 140 languages.

A bilingual text-to-image research model emphasizing text rendering and Chinese semantics; this date records the technical publication.

A 32B open reasoning branch using reinforcement learning to improve complex problem solving.

The image-to-video branch of HunyuanVideo, using a reference image and prompt to guide generation.

A research preview scaling unsupervised learning for knowledge, natural conversation, and creative writing rather than test-time reasoning.

Released open 1.3B and 14B text/image-to-video weights plus VACE general video editing.

Combined near-instant responses and controllable extended thinking in one model, alongside Claude Code.

xAI’s reasoning-agent generation, extending complex-task capability through Think and DeepSearch modes.

StepFun’s open text-to-video release group containing Step-Video-T2V and a faster Turbo variant.

A low-latency multimodal model that introduced native tool use and the Flash Thinking experimental path.

A small text reasoning model for science, mathematics, and coding with selectable reasoning effort; it does not accept images.

A vision-language release group with 3B, 7B, and 72B tiers for image, document, and video understanding.

Used large-scale reinforcement learning to shape reasoning and released both full and distilled model weights.

Moonshot’s multimodal reasoning work exploring long-context reinforcement learning and test-time reasoning scale.

Luma’s second-generation large video model with stronger natural motion, physics, and camera expression.

A hybrid MoE using Lightning Attention, with 456B total and 45.9B active parameters, trained at 1M context and extrapolatable to 4M.

The vision-preview branch of Moonshot V1, accepting images and recorded separately from the text-only V1 tiers.

A 671B-total, 37B-active MoE model that extended the efficient-training and open-release path.

Advanced Google’s video line with stronger physics, cinematic language, and research output up to 4K.

Sora’s first production release with up to 20 seconds, 1080p, extension, blending, and storyboards.

Brought near-Llama-3.1-405B instruction quality to a much smaller 70B model.

The full o1 release, bringing stronger reasoning and image understanding to ChatGPT.

Tencent’s first open Hunyuan video model, with over 13B parameters and released weights and inference code.

A major update to prompt response, motion, and image quality; the official retrospective dates its launch to December 2024.

A representative coding branch of Qwen2.5, with 32.5B parameters and Base and Instruct weights.

Tencent’s first large open MoE with 389B total and 52B active parameters.

The fast successor to Claude 3 Haiku, with stronger coding, instruction following, and tool use.

The 8.1B base model in Stable Diffusion 3.5 for quality, prompt adherence, and professional customization.

The 11B and 90B Llama 3.2 vision tiers for image understanding, separate from the lightweight text-only branch.

The lightweight text-only Llama 3.2 tier for mobile and edge devices, separate from the 11B/90B vision models.

The largest general-purpose dense Qwen2.5 tier, improving coding, mathematics, and structured output; the 72B tier is not Apache-licensed.

An early OpenAI reasoning preview designed to spend more computation before answering complex questions.

A smaller reasoning model released alongside o1-preview, focused on coding and mathematics.

Adobe’s text- and image-to-video model, demonstrated in 2024, public beta in February 2025, and generally available in April.

Combines the general and coding capabilities of DeepSeek-V2-0628 and Coder-V2-0724 in a unified chat model.

MiniMax’s first video-generation model, supporting text- and image-to-video at 720p, 25fps, and up to six seconds.

A major Google text-to-image generation with better detail, lighting, and prompt adherence.

Launched in beta on X with improved chat, coding, and reasoning; product image generation used the separate FLUX model.

BFL’s first 12B open-weight developer model under a non-commercial license; hosted Pro and Apache-2.0 Schnell are not treated as aliases.

Improved image coherence, detail, and texture while accelerating generation; it served as Midjourney’s default model.

Llama’s first 405B open flagship, extending the family to 128K context.

A text-and-vision model for inexpensive, low-latency calls, launched with 128K context and text output.

StepFun’s second-generation general model, confirmed by the official release retrospective as a formal Step-line release on 2024-07-04.

A vision-focused Step upgrade listed in the official release retrospective as a multimodal model released on 2024-07-04.

The 27B open Gemma 2 checkpoint, improving reasoning quality while targeting accessible single-accelerator deployment.

A major reasoning, vision, and coding upgrade at the Sonnet cost and speed tier, updated again in October 2024.

Runway’s video foundation generation improving motion, detail, and temporal consistency; the first announcement is distinct from later product access.

A coding MoE continued from an intermediate DeepSeek-V2 checkpoint, with 236B total and 21B active parameters and 128K context.

Introduced MMDiT in a 2B model aimed at consumer GPUs.

The largest dense Qwen2 model, using GQA across the family and extending support to 27 additional languages.

The public starting point of Kling video generation, initially offering access by application before expanding to global web users.

The efficient Gemini 1.5 tier for high-throughput multimodal work with a million-token context window.

Google’s original Veo video model, introduced at I/O with a private VideoFX preview for selected creators.

An “omni” model natively spanning text, vision, and audio with substantially lower interaction latency.

Combined Multi-head Latent Attention with fine-grained MoE routing to reduce training and inference cost.

Introduced in beta in Firefly, Photoshop, and InDesign, improving image quality, style control, and creative exploration.

Represents the separate 8B and 70B checkpoints released together in April 2024, without assigning one parameter count to the whole release.

The preview announcement of Grok’s first vision model, extending understanding to images, charts, and documents.

Expanded context to 128K and improved mathematics and coding, announced ahead of access for early testers.

Grok-1’s raw pretrained base weights: a 314B MoE with roughly 25% active per token, not the Grok chat model.

The fastest and lowest-cost Claude 3 model for high-throughput, low-latency workloads.

The flagship Claude 3 model, combining visual input with a 200K context window.

The balanced Claude 3 tier, designed for enterprise-scale use across intelligence, speed, and cost.

Represents the separate 2B and 7B first-generation Gemma checkpoints without assigning one parameter count to the whole release.

Used an MoE architecture to extend generally available context to 2M tokens while retaining cross-modal retrieval.

The original Sora research demonstration, exploring video generation with spacetime patches; it was not yet a public generation service.

The 72B Qwen1.5 tier with improved multilingual capability and a 32K context window.

The first formal model series on the Kimi Open Platform, offered in 8K, 32K, and 128K context sizes and powering the Kimi assistant and API.

A 16.4B MoE using fine-grained expert segmentation and shared-expert isolation, released in Base and Chat variants.

Z.ai’s 2024 foundation-model generation advancing language, reasoning, coding, and tool use together.

Significantly improved long-prompt understanding, coherence, knowledge, and image prompting.

An image-generation model offered to approved Vertex AI customers, improving image quality and text rendering.

The first Gemini flagship, designed from training onward for native text, image, audio, and video understanding.

The first Gemini generation’s Pro tier for broad product deployment, initially introduced in Bard.

StepFun’s early native multimodal model, identified in the official retrospective as the visual starting point of the Step line released in late 2023.

DeepSeek’s first general language-model line, with 7B and 67B Base and Chat variants; the official V2 material refers to the 67B base as DeepSeek-V1.

Expanded context to 200K, improved long-document understanding, and introduced system prompts and beta tool use.

Entered developer preview with a 128K context window, expanding GPT-4’s long-document capabilities.

The model behind the original limited Grok beta. The later open release contains base weights without dialogue fine-tuning and is listed separately.

DeepSeek’s earliest public coding line, spanning 1B–33B sizes and adding fill-in-the-middle (FIM) training for project-level completion.

The third ChatGLM dialogue generation with stronger tool-use and code-execution support; this is the standard 8K chat model, not the separate 32K/128K versions.

Baidu’s fourth-generation foundation model, improving understanding, generation, reasoning, and memory with invited product testing.

The second Firefly image model, introduced in web-app beta with improved image quality and creative control.

The initial public enterprise Hunyuan product, introduced for testing and API access through Tencent Cloud.

A coding-specialized model continued from Llama 2; 34B was the largest launch tier, with 70B added later.

The starting point of Qwen’s open-model path, emphasizing Chinese-English, multilingual, and tool-use capabilities.

Used a base-plus-refiner pipeline to bring open generation to native 1024 resolution.

The 70B tier of Llama 2, available for research and commercial use under its community license.

Expanded to a 100K context window and reached broader users through the API and claude.ai.

The second-generation bilingual dialogue model, using multi-query attention and expanding context to 32K.

The successor family to PaLM, improving multilingual, reasoning, and coding capabilities and powering Bard and other Google products.

Firefly’s first image model, generating images and text effects and beginning Adobe’s generative creative-product line.

The first public ERNIE Bot product milestone, offering invited access and applications for enterprise API testing.

A large multimodal model for complex professional tasks; core scale and training details were not disclosed.

Anthropic’s first public Claude model, centered on its Constitutional AI training approach.

A 6.2B open model optimized for Chinese-English question answering and dialogue, and the public starting point of the ChatGLM product line.

The GPT-3.5 model for the message-based chat API, initially released as the gpt-3.5-turbo-0301 snapshot.

A historical version emphasizing photographic output and prompt matching; release month is separate from default-version changes.

The 65B tier of the first LLaMA generation, initially offered by application for noncommercial research.

Transforms input video using text or image guidance; this date records the month on the official research page.

A refined SD 2.0 checkpoint with adjusted training-data filtering and improved prompt response and composition.

ChatGPT’s original dialogue model, fine-tuned from GPT-3.5 with human feedback and introduced as a free research preview.

A diffusion generation using an OpenCLIP text encoder, initially offered in 512px and 768px image tiers.

A generation with a new codebase and architecture, improving detail, complex scenes, and image prompts.

Brought strong text-to-image weights to local GPUs and catalyzed a broad tooling and fine-tuning ecosystem.

An early 130B bilingual open model in the GLM family, trained with the General Language Model paradigm.

A text-to-image research model combining a frozen language encoder with cascaded diffusion; this record marks its paper release.

A 540B dense language model trained with Pathways, demonstrating large-scale multitask and few-shot capabilities.

An early default Midjourney image model with an abstract, painterly style; official history gives month precision.

GPT-3 models aligned with human feedback to follow instructions; this node represents a series, not a single parameter tier.

Google’s language-model research line for open-ended dialogue, focused on conversational quality and topic development.

A 175B autoregressive model that systematically demonstrated few-shot and in-context learning without fine-tuning.

Demonstrated broad task transfer from unsupervised language modeling and introduced a staged-weight release.

A Transformer using generative pretraining followed by task-specific fine-tuning, starting the GPT line.

The high-speed API serving variant of GLM-5.3-Flash, retaining the Flash series’ visual understanding, tool use, and agent capabilities.

The first Grok Voice Transcribe API model, now being replaced by 2.0.

The older high-quality image API model, scheduled for retirement on 2026-11-02 and redirecting to Image 2.0 at low quality.

StepFun’s current audio product group spanning ASR, realtime voice interaction, voice generation, TTS, and music generation.