LLM Models

Reve AI's June 2026 image model, introducing a planning-then-rendering architecture that builds code-based intermediate layouts for native 4K output, lossless iterative editing, and precise composition control. Reve's typography engine (trained on 50M font samples) targets ~98% in-image text accuracy.

Alibaba's open-source 20B-parameter image foundation model (August 2025), an MMDiT architecture that ranks first across nine public benchmarks including GenEval, DPG, and OneIG-Bench. It leads on complex text rendering (Chinese and English at commercial quality) and precise image editing.

ByteDance's unified image generation and editing model (2025), combining creation and editing in a single architecture. Seedream 4.0 produces up to 4K images with ~1.8-second generation, strong reference consistency, and batch processing, and is served via API partners like fal and Replicate.

Midjourney's flagship image model (April 2025, default since June 2025), prized for its distinctive, highly aesthetic output. V7 adds voice prompting, a personalization model, and a Draft mode that renders at ~10x speed and half the cost, with richer textures and more coherent hands and objects.

Recraft's October 2024 design-focused image model (codenamed 'red_panda'), which topped the Artificial Analysis Image Arena at launch. It excels at brand-consistent design, precise text placement, and is one of the few models to generate editable SVG vector graphics.

Ideogram's March 2025 text-to-image model, the specialist for accurate in-image text and typography, with notably improved photorealism over 2.0. It offers style references, magic prompt enhancement, and precise layout control for posters, logos, and designs.

DISCONTINUED. Google DeepMind's flagship text-to-image model, delivering top-tier photorealism, improved text rendering, and strong prompt adherence. Imagen 4 is available through the Gemini API and Vertex AI, with an Ultra tier for the highest-fidelity generations. Google shut down the Imagen 4 standard, ultra and fast endpoints on August 17, 2026; the models index lists it as deprecated.

Stability AI's 8-billion-parameter open text-to-image model (October 2024), the largest of the SD 3.5 family. Built on the Multimodal Diffusion Transformer (MMDiT) architecture, it improves prompt adherence and image quality over SDXL and is released under the permissive Stability Community License.

The open-weight member of Black Forest Labs' FLUX.2 family (November 2025), released on Hugging Face for self-hosting and fine-tuning. FLUX.2 [dev] brings the family's 32B-parameter photorealism and prompt adherence to local and customizable image pipelines.

Black Forest Labs' commercial flagship text-to-image model (November 2025), part of the 32-billion-parameter FLUX.2 family. FLUX.2 [pro] delivers state-of-the-art photorealism and prompt adherence via API, alongside the open-weight FLUX.2 [dev] and the sub-second FLUX.2 [klein] variants.

Black Forest Labs' 12B parameter rectified flow transformer for text-to-image generation. Created by ex-Stability AI developers, released August 2024. Three variants: schnell (Apache 2.0, 1-4 steps, fast), dev (guidance distillation, efficient), pro (state-of-the-art quality). Uses Diffusion Transformer combining Transformer and CNN architectures. Supports resolutions: 1024x1024, 768x1344, 1344x768, 1216x832. Schnell variant generates high-quality images in only 1-4 steps using latent adversarial diffusion distillation. Enables commercial use (schnell/pro), dev for non-commercial.

DISCONTINUED. Mistral AI's first code-specialized open-weight model with 22B parameters. Released May 2024, designed explicitly for code generation tasks including function completion, test writing, and fill-in-the-middle code completion. Trained on diverse dataset of 80+ programming languages (Python, Java, C, C++, JavaScript, Bash, Swift, Fortran, etc.). Features 32K context window - larger than competitors' 4K-16K. Licensed under Mistral AI Non-Production License for research/testing; commercial licenses available on demand. Helps developers write and interact with code through unified instruction and completion API. Mistral deprecated the 24.05 hosted model on December 2, 2024 and retired it on June 16, 2025; a current Codestral (v25.08) exists under Mistral's Premier licence, so this record covers the retired original.