Model releases.

Generates video from text or images with realistic physics and strong visual fidelity.

Generates images quickly, optimized for low-latency creative output.

Generates photorealistic and artistic images with professional-grade quality.

A fully open 32B language model with open weights, training data, and recipe.

A fully open 7B text model with open weights, training data, and recipe.

Generates images with accurate text, multi-image blending and studio composition at 4K.

Converts English text into speech on-device with ten preset voices and no API key needed.

Generates custom sound effects from text prompts.

A fast, cost-efficient text model built for strong reasoning, coding, and tool use.

Generates high-quality video with native audio and frame-to-frame control.

A fast video model that generates high quality short clips from text prompts.

Converts up to 1,000 characters of text into spoken audio using English or Chinese voices.

Fast image generation from one to three tagged references for consistent results.

Keeps a character from one reference picture consistent across new scenes and styles.

A 20B open reasoning model with adjustable effort and tool calling, runs on 16 GB.

Generates images with a 20B model, rendering long English and Chinese text, fully offline.

Generates 3 to 10 second sound effect clips from a short text description.

Turns text prompts into video using a two-expert model that switches during denoising.

Generates video from text or images, running fully offline with no API key needed.

Handles complex clinical reasoning and biomedical analysis as a large medical text model.

A 24B instruction-tuned model with vision input, function calling, and 128K context.

A reasoning model for math, code, and logic, distilled from Qwen3 into 8B parameters.

A 14B model fine-tuned for chain-of-thought reasoning on math, science, and code.

Generates images with strong prompt adherence and consistent characters via tag refs.

An ultra-compact 0.6B model with tool use that runs in-process or through Ollama.

A compact 4B instruction model with strong reasoning and multilingual capabilities.

Turns a starting image into a 2 to 10 second video clip, quickly and cheaply.

Converts text to speech with ten natural voices steerable by free-text instructions.

A compact 1B model that runs in-process or through Ollama.

A compact 3.8B text model with a 128K context window and native function calling.

A lightweight local video model that generates 480p clips on modest hardware.

A compact 1.5B reasoning model for on-device chain-of-thought, distilled from Qwen2.5.

A 14B reasoning model distilled from Qwen2.5, excelling at math and coding benchmarks.

A 32B reasoning model distilled for top-tier chain-of-thought performance.

An English-only speech model with very low latency for real-time use.

The fastest speech model, about 75 ms latency across 32 languages at a lower price.

A 14B open model built for strong reasoning, math, and code tasks.

Fast, slimmed-down speech-to-text model that keeps accuracy and returns timed subtitles.

The smallest Llama, a 1B instruction-tuned model that runs in-process or through Ollama.

Runs fast on-device text generation using a compact 3B instruction-tuned model.

Handles general reasoning and instructions with a 128K token context window.

Transcribes speech in nearly a hundred languages and returns timed subtitle segments.

A voice model for consistent long-form narration and audiobooks, in 29 languages.

Transcribes speech on-device with no API key, favoring speed over accuracy.

Transcribes speech on-device with no API key needed, balancing speed and accuracy.