Model releases.

Answers yes/no, choice and score questions with calibrated probabilities, 262K context.

Transcribes recorded audio with word timings, speaker labels and key-term hints.

Google's most capable reasoning model for hard coding, science and agents, 1M context.

Generates and edits images from up to 16 references, at up to 20 megapixels.

A faster H3 Image tier that generates and edits images from up to 16 references.

Handles the most demanding reasoning and long agentic work, with vision and 1M context.

A small, efficient Muse model for tool calls, routing and extraction, with a 1M-token context.

A finance-tuned Drex Max decision model, with a 262K-token context.

Generates, edits and extends video up to 15 s with synced audio, from 360p to 4K.

Investigates vulnerabilities and verifies fixes, for approved security teams.

Faster 720p Kling 4.0 clips up to 20 s with audio, from text, a frame or references.

Fast image generation and editing, with vector SVG and animated GIF output.

Detailed images and consistent edits, with vector SVG and animated GIF output.

Generates foley, ambiences and impacts up to 30 seconds from several cues at once.

Biological reasoning and scientific tool use, for approved life sciences organizations.

Text-to-speech for voice agents, narration and multilingual apps, with twenty voices.

Transcribes recordings of up to an hour in multiple languages.

An offline decision model for text and images on Apple silicon Macs.

An on-device model that speaks and transcribes 19 languages.

A small on-device model for text and images.

Fast, efficient tier for visual reasoning, computer use and writing, with 1M context.

Works longer on each answer for higher quality results on hard text tasks.

A faster H3 tier for everyday 768P video, with frames and references.

A lower-cost H3 tier for drafts and volume, with frames and references up to 768P.

An open-weight reasoning model for long-horizon coding and agents, with 1M context.

Successor to D1 for choice, score and yes/no answers about text and images.

A small vision-language model for fast local chat, images and tool use.

Generates clean, editable SVG artwork with sharp typography from a prompt.

Traces an attached image into a scalable, editable SVG.

Returns typed, probabilistic answers to choice, score and yes/no questions about a state.

Faster, lower-cost text-to-speech with the same voices, and voice cloning on Runware.

A 122B MoE model with 256K context for reasoning, coding, and agentic tool use.

A tiny on-device English text-to-speech model for Apple silicon Macs.

A lower-cost transcription model with timestamps and language detection.

Mistral's most accurate transcription model, with timestamps and language detection.

Text-to-speech with eight preset voices, directed by emotion and pace tags in the text.

A low-cost coding and agent model with 1M context and tunable thinking.

A fast, low-cost mixture-of-experts model with image input, 1M context and tool calling.

An open 5B text-to-image model that runs offline in 50 steps.

Perplexity's 27B decision model for choice, score and yes/no answers on text and images.

An open-weight reasoning model for coding and long agentic tasks, with 1M context.

A fast, low-cost FLUX 3 model that generates and edits images, up to native 4K.

An offline FLUX 3 model that generates from prompts and edits attached images.

Edits images from plain-language instructions and generates from text, at 1K or 2K.

Video up to 30 s at 1080p with native audio, from text, frames, references or a clip.

Liquid AI's largest decision model, for choice, score and yes/no answers on images too.

Meta's open-weight reasoning model for agentic coding, with image input and 1M context.

A compact vision-language model for coding and tools, sized for a 24 GB NVIDIA GPU.

An open multimodal model for coding, reasoning and documents, with 1M context.

A ternary-weight vision model with tools and thinking, about half a 4-bit model's size.