01 / TIMELINE
Browse major model releases since 2018 and filter by year, organization, family, release type, or weights explicitly stated by each event.
DeepSeek updated its web, app, and API model to V4.1-Flash-1017 while keeping the stable API name `deepseek-flash` and pricing unchanged. This dated version has a separate model repository with MIT-licensed weights. DeepSeek reports improved benchmark performance, faster task completion, and shorter outputs.
Microsoft AI announced that pretraining is complete for MAI-Base-2, a 2.8T-total / 80B-active MoE base model now entering post-training; no user release or public weights are available yet.
BFL released FLUX 3 Image Klein, a 5.6B image generator for text-to-image, editing, and multi-reference workflows, with weights and API access on launch.
Kuaishou launched Kling 4.0 on 2026-10-15. It supports single-pass video generation up to 30 seconds, up to 15 mixed reference assets, and up to 10 keyframes. API access and 4K/1080p output are planned for later.
Meta released Muse Spark Open weights. Its official model card lists BF16 and NVFP4 checkpoints, 816B total / 42B active parameters, and 1M context; the model uses the Muse Model License.
StepFun opened Step 5 Preview weights on the announced date; FP8, BF16, and NVFP4 checkpoints are accessible under its official organization, with the StepFun Model License 1.0. The model remains in Preview.
OpenAI is rolling out GPT-Live-1.1 and the lighter Quip in ChatGPT Voice; its system card identifies them as successors to GPT-Live-1 and GPT-Live-1 mini. Both retain full-duplex voice interaction, and these sources publish no API model IDs.
xAI released the real-time voice model `grok-voice-3.0`; `grok-voice-latest` now points to 3.0, which supports speech-to-speech interaction. xAI says the API and price are unchanged, and the previous `grok-voice-2.0` is deprecated.
Gemini 4 Argon became globally available to developers and enterprises through the API on October 13, and to Google AI Pro / Ultra users in the Gemini app and Search AI Mode; its API model ID is `gemini-4-argon`.
Meta released two distinct models on the same day: Muse Wonder (Meta Model API Preview / Muse Code) and Muse Wonder Full, available only to Muse Code subscribers.
Z.ai first introduced GLM-Legal on October 12: it is built on GLM-5.3-Flash with legal-domain mid-training and reinforcement-learning post-training, and API beta applications are open; AutoClaw availability is still forthcoming. The originator has not published a distinct API model ID, parameter count, weights, or license.
Qwen-Image-2.1-Turbo launched with public weights: an 8-step accelerated checkpoint for image generation and editing under a non-commercial research license.
OpenAI introduced the October ChatGPT versions of GPT-6 Sol and Luna. Its system card distinguishes them from the September versions still used in ChatGPT Work/Codex; the new versions add Intelligent UI and updated safety training, with Sol available to paid plans from October 7 and Luna to Free/Go from October 8.
Anthropic launched Claude Haiku 5.5 with model ID `claude-haiku-5-5`; it accepts text and image input, returns text, and supports a 1M context window and 128K output.
Nano Banana 2.1 launched as a stable Gemini API model, updating Nano Banana 2 for image generation and editing.
QwenCloud released the Qwen-Image-2.1-Pro image-generation and editing API, with native transparency and up to 10 reference images.
Microsoft launched MAI-Transcribe-2-Streaming for real-time transcription, along with the multilingual MAI-Voice-2.1 and its low-latency Flash variant.
Google first announced Gemini 4 Argon on 2026-09-30; at the time of this launch post, access was limited to trusted cyber defenders through the Fairwind Program, with broader API and Google AI Ultra access still forthcoming. The post describes chart and long-video understanding and states a 1M-token maximum output.
OpenAI launched GPT-6.1 Sol (`gpt-6.1-sol`) as an upgrade to GPT-6 Sol for coding, computer use and professional work, with text/image input, 1.05M context and 128K output, available in the API, ChatGPT Work and Codex.
Anthropic launched Claude Sonnet 5.5 as the second model in the Claude 5.5 family; API model ID `claude-sonnet-5-5`, available on Claude, AWS, Google Cloud, and Microsoft Azure.
Gemini 3.8 Live gained Live Avatar on the existing `gemini-3.8-live` model, adding real-time voice/video interaction and MP4 avatar-video output; the feature is available in Gemini Enterprise.
QwenCloud released the decision-model-preview API for structured classification, decisions, and scoring rather than text generation.
OpenAI shut down the Sora 2 and Sora 2 Pro APIs and the Videos API; the service is now retired.
Meta Research introduced Muse Realtime Avatar, an audio-driven model that continuously generates synchronized character video from speech tokens and reference media; reported demo specs are 448×768 at 25 fps with about 870 ms to first byte.
Anthropic launched Claude Opus 5.5 as the first model in the Claude 5.5 family for long-running agentic coding and knowledge work, available through the API, AWS, Google Cloud, and Microsoft Foundry.
Gemini 3.8 Flash TTS and Flash-Lite TTS reached GA together in the Gemini API, targeting creative speech and high-throughput, low-latency speech tasks.
OpenAI released GPT-6 Sol and GPT-6 Luna for complex work and high-volume tasks, with text/image input, 1.05M context and 128K output, available through the API, ChatGPT Work and Codex.
Tencent introduced Hy Image 3.5 Preview for image generation and editing, with up to five references, multi-turn creation and 2K output; it remains a preview.
Grok 4.7 launched in the xAI API with a 500K context window, text/image input, controllable reasoning effort, and long-horizon coding-agent work.
QwenCloud released qwen3.8-omni-flash-realtime for real-time audio/video interaction with text/audio output, multichannel audio, video aggregation, remote MCP tools, and WebSocket, WebRTC, and AOQ access.
Xiaomi launched the native omni-modal MiMo-V2.6-Pro and MiMo-V2.6-Flash models with API access and open weights.
QwenCloud released qwen-audio-3.1-realtime-plus for full-duplex realtime voice with text/audio output, function calling, web search, and a 262,144-token context.
Step 5 Preview entered product and API preview on StepFun; open weights are not yet released.
QwenCloud released qwen3.8-omni-flash: a native omni-modal model with up to 1M context, text/image/audio/video inputs, and DashScope/OpenAI protocol compatibility.
QwenCloud released the HappyOyster 1.0 interactive world-model series in Adventure, Directing and Acting modes.
SpaceXAI made Grok Voice Transcribe 2.0 available in the Speech-to-Text API on September 17 and announced it would become the default transcription model; Grok Voice Transcribe 1.0 entered a deprecation transition. The launch article is dated September 18.
QwenCloud released qwen3.8-livetranslate-flash-realtime with audio/image input, text/audio output, 60 source languages, 29 audio-output languages, and WebSocket Realtime API access.
Gemini 3.8 Live and Extended Thinking reached GA together in the Live API, offering real-time voice interaction and background reasoning.
The Z.ai / AutoClaw page’s structured metadata dates GLM-5.3-Flash to 2026-09-15 and confirms MIT-licensed weights with native text, image, video, and file support.
Kimi K2.8 Preview fully rolled out in Kimi Code; the Model ID `kimi-for-coding` is unchanged, with low, high, and max thinking and up to 1M context.
DeepSeek V4.1 Flash reached general release and entered the API under the recommended name `deepseek-flash`; the old V4 Flash and Vision Exp models were retired and temporarily route to the new model.
OpenAI launched GPT-Live-1 in the API with full-duplex voice conversation and delegation of deeper reasoning and tool use to a backend model or agent.
GPT Image 2.5 (ChatGPT Images 2.5) launched, with Flare and Sunburst introduced in the API.
Xiaomi opened invitation-based desktop testing for MiMo-X-Pro-Preview and MiMo-X-Flash-Preview.
MAI-Image-2.6-Flash entered preview.
GPT-6 Astra launched and expanded from Trusted Access into a phased rollout across Plus, Pro, Business, and Enterprise plans; Enterprise/Edu access remains controlled by workspace permissions.
Muse Spark 1.3 officially launched and began rolling out in Muse Code and Meta Model API.
Gemini 3.8 Flash launched with a 1M input context, 65K output limit, multimodal inputs, and thinking capabilities.
Gemini 3.8 Flash Cyber became available to governments and trusted partners through the limited-access Fairwind Program.
Qwen3.8-Max-0902 launched as an upgraded Qwen3.8-Max snapshot, improving coding, cowork, long-horizon autonomous development, collaborative agents, and vision.
Claude Fable 5.1 reached general availability; the same underlying model is offered as Claude Mythos 5.1 through trusted access programs.
Agentic video understanding launched for Gemini 3.7 Flash and rolled out across Gemini 3.6 Flash and 3.5 Flash-Lite, reducing long-video analysis by up to 88% in tokens and 66% in cost.
Muse Voice Transcribe launched in Meta Model API, Meta AI for Mac, and Muse Code.
GLM-5.3 weights were released with FP8 and BF16 checkpoints under the custom GLM-5.3 License.
Gemini Omni 1.1 Flash reached GA with video extension, first/last-frame interpolation, and resolution control.
Qwen3.8-Flash-Next released open weights and became available as managed Qwen3.8-Flash.
DeepSeek-V4-Flash-Vision-Exp launched in the multimodal API.
GLM-5.3 launched with gains attributed entirely to scaled post-training.
DeepSeek-V4-Pro-0813 reached general availability with adjustable reasoning effort.
Gemini 3.7 Flash reached general availability.
Grok 4.6 launched for long-running agents and complex visual work.
OpenAI released GPT-5.6-Cyber; the API model is built on GPT-5.6 Sol and available to defenders through approval-gated Daybreak Red.
MAI-Image-2.6 entered preview.
Grok Imagine Image 2.0 launched for image generation and editing through the API.
DeepSeek-V4-Flash-0731 entered public beta in the API.
Gemini 3.5 Flash-Lite reached general availability.
Gemini 3.6 Flash reached general availability.
Kimi K3 was announced at 2.8T parameters with native vision.
GPT-5.6 reached general availability across ChatGPT, Codex, and API.
Muse Spark 1.1 entered public preview through Meta Model API.
Claude Sonnet 5 launched and became the default for Free and Pro.
Gemini Omni Flash Preview entered the video-generation API.
Nano Banana 2 Lite launched.
Kimi K2.7 Code launched as an open model for long-horizon coding and agent tasks.
Claude Fable 5 reached general availability alongside Mythos 5.
Microsoft introduced MAI-Thinking-1 at Build, ahead of its August Foundry public preview.
Step 3.7 Flash launched with open weights and deployment materials.
Microsoft announced the MAI-Image-2.5 evaluation release on Arena.
Google announced Gemini 3.5 Flash at I/O.
Gemini 3.1 Flash-Lite reached general availability.
GPT-5.5 Instant updated ChatGPT’s default fast experience.
Gemini Omni Flash Preview was announced.
Step Image Edit 2 launched as a low-latency image generation and editing service.
The Step-Audio-R1.5 technical report and evaluation benchmarks were published.
Sora web and app were discontinued while the API entered its shutdown window.
DeepSeek released public previews and open weights for V4 Pro and Flash together.
Xiaomi opened public beta testing for MiMo-V2.5 and MiMo-V2.5-Pro together; the later weight release is recorded separately.
GPT Image 2 and ChatGPT Images 2.0 launched.
Claude Opus 4.7 reached general availability.
GPT-5.4-Cyber launched as a cybersecurity fine-tune of GPT-5.4, with limited access for vetted defenders through Trusted Access for Cyber.
Google released the open Gemma 4 31B and 26B MoE models together.
GPT-5.4 mini launched for coding and subagents.
GPT-5.4 Thinking and Pro launched in ChatGPT, API, and Codex.
GPT-5.3 Instant became ChatGPT’s everyday fast model and entered the API.
GLM-5 launched for complex systems engineering and long-horizon agents.
GPT-5.3-Codex-Spark launched as a real-time coding research preview.
GPT-5.3-Codex launched across Codex products.
Kuaishou launched the Kling AI 3.0 multimodal line.
NextStep-1.1 launched with pretraining and post-training weights.
GPT-5.2-Codex launched for long-horizon software engineering and defensive security.
Kling Video 2.6 launched with synchronized audio and video generation.
DeepSeek released open weights for V3.2 and its high-reasoning Speciale model.
Kling O1 Video launched.
Claude Opus 4.5 launched with effort control.
Grok 4.1 Fast and the Agent Tools API launched.
GPT-5.1 launched in the API with adaptive reasoning and new developer tools.
ERNIE 5.0 was announced.
MAI-Image-1 began rolling out in Bing Image Creator and Copilot products.
Gemini 2.5 Flash Image launched as Nano Banana.
The NextStep-1-Large series released inference code and image-model weights.
OpenAI launched GPT-5 alongside its smaller mini and nano API models.
OpenAI released the gpt-oss-120b and gpt-oss-20b weights together.
Step-Audio 2 technical report and audio-model line were introduced.
Kontext dev released its 12B weights under the FLUX.1 Non-Commercial License.
Midjourney launched its first image-to-video model.
Gemini 2.5 Flash became generally available.
Gemini 2.5 Pro became generally available.
The Kontext family was announced and dev entered private beta; weights were not yet public.
Kling 2.1 launched.
DeepSeek-R1-0528 launched with stronger reasoning, instruction following, and tool use.
Anthropic launched Claude Opus 4 and Claude Sonnet 4 together.
Kimi Thinking Preview entered preview.
Adobe launched the Firefly Image 4 and Ultra image models together.
Firefly Video Model became generally available.
GPT Image 1 entered the API.
Gemini 2.5 Flash entered preview.
OpenAI launched the o3 and o4-mini reasoning models together.
Kling 2.0 launched.
GPT-4.1, GPT-4.1 mini and GPT-4.1 nano launched together in the API.
The Step-R1-V-Mini multimodal reasoning model page was published.
Meta released the open-weight Llama 4 Scout and Maverick multimodal MoE models.
Gemini 2.5 Pro Experimental launched with thinking built into the new generation.
The deepseek-chat API upgraded to DeepSeek-V3-0324 with improved reasoning, coding, and tool use; the standalone announcement followed on March 25.
Step-Video-TI2V inference code, weights, and technical report were released.
Claude 3.7 Sonnet launched as a hybrid-reasoning model.
Grok 3 Beta was unveiled as a reasoning-agent generation.
Step-Video-T2V and T2V-Turbo inference code and weights were released.
Firefly Video Model entered public beta.
Gemini 2.0 Flash became generally available through the API and Vertex AI.
DeepSeek-R1 and its distilled models were released with open weights.
Moonshot published its Kimi K1.5 multimodal reasoning work.
Moonshot V1 Vision Preview entered preview.
Anthropic announced Claude 3.5 Haiku.
An upgraded Claude 3.5 Sonnet launched with computer use in public beta.
OpenAI released the o1-preview and o1-mini reasoning-model previews.
Firefly Video Model was announced.
Kling opened its web beta to users worldwide.
GPT-4o mini became available through an API.
The formal Step-2 release entered StepFun’s public model lineup.
Gemini 1.5 Pro’s 2M context window reached general availability.
The deepseek-coder API upgraded to DeepSeek-Coder-V2-0614, ahead of the open-weight release.
Gemini 1.5 Flash reached general availability.
Firefly Image 3 entered preview.
Anthropic released Claude 3 Opus and Sonnet; Haiku became available later.
Gemini 1.5 Pro entered limited preview with a 1M context window.
Sora (Research) was announced.
The Kimi Open Platform entered public beta with the Moonshot V1 series supporting up to 128K context.
DeepSeekMoE 16B was released publicly, exploring efficient MoE through fine-grained and shared-expert routing.
Z.ai released the GLM-4 foundation-model generation.
Google announced the natively multimodal Gemini 1.0 family.
Gemini 1.0 Pro launched.
Step1V was released at the end of 2023 as an early native multimodal Step milestone.
DeepSeek LLM 7B/67B Base and Chat models were released, forming the DeepSeek-V1 baseline referenced by the official V2 material.
GPT-4 Turbo entered preview.
The DeepSeek-Coder line was first released publicly across several sizes for code completion.
Firefly Image 2 entered preview.
Hunyuan (2023) became available through an API.
ChatGLM launched with the open ChatGLM-6B version optimized for Chinese-English dialogue.
GPT-3.5 Turbo became available through an API.
GLM-130B was released with open weights as an early large-scale public model in the GLM family.
Imagen was announced.
InstructGPT became available through an API.
The GPT-3 paper documented few-shot learning at 175B parameters.
OpenAI announced GPT-2 and began a staged release.