AI Models & Foundation Models

OpenAI's flagship multimodal model. Fast, intelligent, and natively handles text, images, and audio in a single model. Powers ChatGPT Plus.

GPT-4 with 128K context window and April 2024 knowledge cutoff. More capable than GPT-3.5, optimized for cost.

Anthropic's most intelligent model. Best-in-class at coding, research, and nuanced reasoning. Features computer use capability.

Anthropic's most powerful model for highly complex tasks. Top-level performance on evaluation benchmarks including expert-level tasks.

Fastest and most compact Claude model. Near-instant responses for lightweight tasks. Excellent cost-to-performance ratio for production.

Google's breakthrough model with 1M token context window — can process an entire codebase, long videos, or books in one prompt.

Google's fastest Gemini model with real-time streaming, tool use, and multimodal output generation.

Meta's largest open-source model. Rivals GPT-4 and Claude 3 Opus on benchmarks. Fully open weights — run it locally or on any cloud.

Meta's multimodal Llama with vision capabilities. Includes small 1B/3B models for edge devices and 11B/90B for powerful vision-language tasks.

Mistral's flagship model. Top-tier reasoning, coding, and multilingual performance. European AI independence with Apache 2.0 licensing.

Mistral's open MoE model. Uses 8 expert networks, activating 2 per token. Efficient, powerful, Apache 2.0 licensed.

Alibaba's powerful open model. Excellent at coding, math, and multilingual tasks — particularly strong in Chinese.

DeepSeek's frontier MoE model trained on 14.8T tokens. Top-tier coding and reasoning at a reported $6M training cost — disrupted the industry.

xAI's most powerful model. Trained on massive compute cluster. Features "Think" mode for extended reasoning and real-time X/Twitter data access.

Cohere's enterprise LLM optimized for RAG and tool use. Excellent at grounded generation with citations.

Google's open-weight model series. Gemma 2 27B matches or beats much larger models. Fully open, commercially usable.

UAE's open-source LLM trained on 3.5T tokens. Was briefly the top open-source model. Apache 2.0 licensed.

Strong bilingual (English + Chinese) open model with exceptional 200K context window. Excels at long-document tasks.

Korean startup's model using depth up-scaling (DUS). Outperformed many 70B models at 10.7B scale.

OpenAI's most powerful reasoning model. Uses extended chain-of-thought to solve complex math, science, and coding problems with near-human accuracy.

First of OpenAI's reasoning series. "Thinks" before answering using internal chain-of-thought. Dramatically better at STEM problems.

Chinese open-source reasoning model that shocked the world — rivals o1 at a fraction of the training cost. Full weights released under MIT license.

Microsoft's tiny but mighty 3.8B model. Punches well above its weight class for reasoning, math, and coding. Runs on phone CPUs.

Third generation of the landmark open-source image model. Uses Multimodal Diffusion Transformer (MMDiT) for better text rendering and prompt adherence.

Stable Diffusion XL — major upgrade with a two-stage pipeline. Better faces, photorealism, and composition at 1024×1024.

State-of-the-art open image model from ex-Stability AI team. Outperforms Midjourney and DALL-E 3 in quality benchmarks. Apache 2.0 licensed.

FLUX.1 Schnell (German: "fast") — 4x faster than FLUX.1 dev with only slight quality reduction. Apache 2.0, fully commercial.

Google's highest quality image model. Best photorealism and text rendering in Google's lineup. Available via Gemini and Vertex AI.

OpenAI's most capable image model. Built into ChatGPT. Exceptional at following long, complex prompts and generating readable text in images.

Industry gold standard for artistic AI images. v6 adds natural language prompting, improved realism, and coherent text generation.

Best-in-class AI image model for text typography and logos. v2 dramatically improves realism while maintaining text accuracy.

Breakthrough control mechanism for diffusion models. Enables precise control via pose estimation, depth maps, edge detection, and other conditions.

Image Prompt Adapter — use any image as a style/content reference for diffusion models. Works with SD1.5, SDXL, and FLUX.

OpenAI's groundbreaking video generation model using Diffusion Transformers (DiT). Generates up to 1080p video up to 1 minute. Simulates physical reality.

Alibaba's open-source video generation model. Best open video model — generates high-quality, temporally consistent videos with excellent motion.

3D causal VAE video model that understands temporal relationships. 3D-Full Attention mechanism captures motion across frames naturally.

Open video model using Asymmetric Diffusion Transformer (AsymmDiT). Exceptional motion fidelity and natural movement.

OpenAI's state-of-the-art speech recognition model. Trained on 680K hours of audio. Best-in-class transcription across 99 languages.

Meta's open music generation model. Generate music from text descriptions or condition on existing melodies. Single-stage autoregressive transformer.

Meta's non-autoregressive TTS model using flow matching. Can perform zero-shot voice cloning, noise removal, and style transfer.

Unified audio generation model — generates sound effects, music, and speech from text using a single latent diffusion model.

Large Language and Vision Assistant. Open-source multimodal model connecting CLIP vision encoder to LLaMA. Excellent visual reasoning.

Top open-source vision-language model series. 108B variant matches GPT-4V on many benchmarks. Strong at OCR, charts, and complex visual tasks.

Hugging Face's open multimodal model built on Llama 3.1. Strong document understanding and chart reasoning.

Revolutionary protein structure prediction model. Extends to DNA, RNA, and small molecules. Uses diffusion-based architecture — solved 50-year biology problem.

AI weather forecasting model that outperforms traditional numerical models. Predicts 10-day forecasts in under 60 seconds on a single TPU.

World's largest genomics foundation model. Trained on 9.3T nucleotides across all domains of life. Can design novel proteins and predict mutations.

Segment Anything Model 2 — zero-shot segmentation for images AND video. Click once to track any object through an entire video sequence.

Self-supervised vision foundation model. No labels needed — learns universal visual features. Powers depth estimation, segmentation, and more.

Mistral's specialized code model. Best-in-class at 80+ programming languages. Fill-in-the-middle (FIM) support for IDE integration.

DeepSeek's open-source code model matching GPT-4 Turbo on coding benchmarks. 338 programming languages, MoE architecture for efficiency.

Open-source code model trained on 600+ programming languages from The Stack dataset. Strong FIM, multi-file context, and repo-level reasoning.

OpenAI's best embedding model. 3072 dimensions, 64% cost reduction vs ada-002. Industry standard for RAG, semantic search, and classification.

Best open-source embedding model using instruction tuning on Mistral 7B. Top of MTEB leaderboard among open models.

Reasoning + Acting framework. LLM interleaves thought, action, and observation. Foundation of modern AI agents — used by LangChain, AutoGPT, and more.

Seminal prompting technique. Adding "think step by step" dramatically improves LLM reasoning. Inspired o1, R1, and all reasoning models.