AI Models by Family

Qwen is Alibaba’s open-weight model family. The Qwen series leads many open leaderboards for code, math, and multilingual benchmarks and includes Mixture-of-Experts and dense variants from 0.5B to 235B parameters.

GPT is OpenAI’s flagship language model family. The closed-source GPT-4o, GPT-4.1, GPT-5, and o-series reasoning models are the most widely used commercial AI products in the world, accessed through the OpenAI API and ChatGPT.

Claude is Anthropic’s closed-source frontier model family. The Sonnet, Haiku, and Opus tiers are known for long-context reasoning, agentic tool use, and high-quality writing, accessed through the Anthropic API and Claude.ai.

Gemini is Google’s closed-source frontier model family. Gemini 2.5 Pro and Flash power Vertex AI, the Gemini app, and Google Search’s AI Overviews. Variants span fast multimodal models to long-context reasoning tiers.

Grok is xAI’s closed-source model family, with weights for Grok 1 released under Apache 2.0 and later versions kept proprietary. Grok models are accessed through X and the xAI API.

Llama is Meta’s open-weight language model family. Since the original 2023 release it has become the most widely fine-tuned foundation model in the world, with variants spanning 1B to 405B parameters across dense and Mixture-of-Experts architectures.

Mistral is the open-weight model family from Paris-based Mistral AI, including the Mistral, Mixtral, Codestral, and Devstral lines. Mixtral pioneered cost-effective open Mixture-of-Experts inference.

GLM is Zhipu AI’s open-weight model family, strong on Chinese, code, and agentic benchmarks. Both base and chat variants ship under permissive licenses.

Hunyuan is Tencent’s multi-modal model family spanning language, image, and video. Several Hunyuan checkpoints are open weight under permissive licenses.

Gemma is Google’s open-weight model family, derived from the same research as Gemini. Sizes range from 270M on-device variants to 27B server-class checkpoints, with strong fine-tuning support.

DeepSeek is a Chinese open-weight model family known for cost-efficient Mixture-of-Experts architectures. DeepSeek V3 and the R1 reasoning model match or exceed closed frontier models on coding and math at a fraction of the inference cost.

FLUX is Black Forest Labs’ text-to-image model family. FLUX.1 [dev] and [schnell] are open weight, while FLUX.1 [pro] is a closed-source API tier. FLUX leads many open image benchmarks.

Kimi is Moonshot AI’s long-context model family. The Kimi K2 series ships under permissive open weights and leads many open agent leaderboards on tool-use and reasoning benchmarks.

MiniMax is the Shanghai lab’s flagship language model family, including MiniMax-Text and MiniMax-M1. Recent releases are open weight and compete on long-context and agent benchmarks.

Veo is Google DeepMind’s closed-source text-to-video model family. Veo 2 and Veo 3 power video generation in the Gemini app and Vertex AI.

Kokoro is an open-weight text-to-speech model family with high voice quality at small parameter counts, popular for on-device speech synthesis.

E5 is Microsoft’s open-weight embedding family, including the multilingual-E5 series widely used for retrieval across languages.

Falcon is TII’s open-weight foundation model family from Abu Dhabi, with dense, Mixture-of-Experts, and small on-device variants released under permissive licenses.

Step is Stepfun’s language model family from Shanghai. Step-1 and Step-2 compete at the open frontier in Chinese and English, with Step-Audio variants for speech.

Kandinsky is Sber AI’s open-weight text-to-image and text-to-video diffusion family. Kandinsky 3 and Kandinsky Video are released under permissive licenses for self-hosting.

Voxtral is Mistral AI’s open-weight speech-and-text model family. The Voxtral series handles transcription, translation, and audio understanding through a unified speech-LLM stack.

Granite Speech is IBM’s open-weight speech-to-text family, part of the Granite portfolio. Designed for enterprise transcription with permissive licensing.

Fish Speech is Fish Audio’s open-weight multilingual text-to-speech family with low-latency streaming and voice cloning, widely self-hosted for real-time speech.

OpenVoice is MyShell’s open-weight voice cloning text-to-speech family with tone-color cloning from a short reference clip, MIT licensed for commercial use.

Parler-TTS is Hugging Face’s open-weight high-fidelity text-to-speech family with natural-language style control, released under Apache 2.0.

Jina Embeddings is Jina AI’s open-weight multilingual embedding family, with long-context and multimodal variants used widely in production RAG stacks.

Phi is Microsoft’s small open-weight language model family. Phi models punch well above their parameter count on reasoning benchmarks and are designed for edge and on-device inference.

Stable Diffusion is Stability AI’s open-weight text-to-image model family. SD, SDXL, and SD3 are the most widely self-hosted image generators in the world, with a massive fine-tuning ecosystem.

Sora is OpenAI’s closed-source text-to-video model. It generates up to one-minute high-fidelity clips with physically consistent motion and is accessed through the Sora app and OpenAI API.

Kling is Kuaishou’s closed-source video model family. Kling text-to-video and image-to-video models lead the Video Arena and are accessed through the Kling API.

Runway Gen is the company’s closed-source text-to-video model family, including Gen-3 and Gen-4. Gen models are tuned for filmmaking workflows.

Yi is 01.AI’s bilingual open-weight language model family. Yi releases compete at the open frontier on Chinese and English benchmarks and ship under permissive licenses.

BAGEL is ByteDance’s open-weight unified image generation and editing model, blending diffusion and autoregressive components for prompt-following text-to-image and instruction-based edits.

Wan is Alibaba’s open-weight text-to-video and image-to-video diffusion family. Wan 2.2 ships at consumer-GPU compute budgets while matching closed video generators on motion quality.

Mochi 1 is Genmo’s open-weight text-to-video diffusion model. Released under permissive licensing, it set the bar for open video generation at launch.

LTX-Video is Lightricks’ open-weight diffusion transformer for fast text-to-video and image-to-video generation, tuned for real-time output on a single consumer GPU.

Moonshine is Useful Sensors’ open-weight low-latency speech-to-text family, designed for on-device transcription on phones, laptops, and microcontrollers.

Granite is IBM’s open-weight language model family, with dense and Mixture-of-Experts variants tuned for enterprise workloads, code generation, and retrieval-augmented generation.

CosyVoice is Alibaba’s open-weight text-to-speech family with multilingual support and zero-shot voice cloning, popular for self-hosted streaming TTS.

XTTS is Coqui’s open-weight multilingual voice cloning text-to-speech family. XTTS v2 is widely self-hosted for high-quality speech synthesis in dozens of languages.

MeloTTS is MyShell’s open-weight multilingual real-time text-to-speech family, MIT licensed and tuned for CPU inference on consumer hardware.

WhisperSpeech is Collabora’s fully open text-to-speech family, trained with a Whisper backbone for permissively licensed voice generation across multiple languages.

Voyage Embeddings are Voyage AI’s closed-source retrieval models tuned for RAG. Voyage-3 and voyage-large lead MTEB on multiple retrieval and reranking tasks.