AI21 Labs
Language models and task-specific APIs aimed at enterprise document and reasoning workloads.
Visit AI21 Labs →The models, the labs behind them, and where to actually run them
Picking a model is now three decisions, not one: whose model, open weights or an API, and hosted by whom. This directory keeps those separate — the labs that train frontier models, the open-weight families you can download, the hubs that distribute them, and the providers that will serve any of it for you.
26 listings · 5 categories · updated September 2026Listed alphabetically within each group. This is not a ranking, nothing here is a paid placement, and no listing was solicited. If you represent something listed and a detail is wrong, write to [email protected] and it will be changed or removed.
Advertisement
Browse models and providers
Grouped by what each one gives you: a proprietary API, downloadable weights, distribution, or hosted inference.
Anthropic
Maker of the Claude family, with a research focus on AI safety, interpretability and long-horizon agentic reliability.
Visit Anthropic →Cohere
Enterprise-focused models for retrieval, reranking and generation, with private-deployment options.
Visit Cohere →Google DeepMind
Maker of the Gemini family, with deep integration into Google Cloud and Workspace products.
Visit Google DeepMind →Meta AI
Research lab behind the Llama family of open-weight models.
Visit Meta AI →Mistral AI
European lab publishing both open-weight and commercial models, with a strong efficiency focus.
Visit Mistral AI →OpenAI
Maker of the GPT family, offered through a commercial API and consumer products.
Visit OpenAI →Black Forest Labs
Maker of the FLUX image-generation models, published in both open and commercial variants.
Visit Black Forest Labs →DeepSeek
Open-weight models with a focus on reasoning and code.
Visit DeepSeek →Gemma
Google's open-weight family, built from Gemini research and sized for single-GPU and on-device use.
Visit Gemma →Llama
Meta's open-weight model family, commonly used as a starting point for self-hosted deployments.
Visit Llama →Phi
Microsoft's small-model family, focused on getting useful capability into very small parameter counts.
Visit Phi →Qwen
Alibaba's open-weight family, notable for strong multilingual coverage and a wide range of model sizes.
Visit Qwen →Stability AI
Open generative models for images, video and audio, widely used as the base for fine-tunes.
Visit Stability AI →Hugging Face
A central hub for open models, datasets and demos, and the usual distribution point for weights.
Visit Hugging Face →Kaggle Models
Model distribution alongside Kaggle's datasets and notebooks, convenient for experimentation.
Visit Kaggle Models →Ollama
The simplest way to pull and run open models locally, with a consistent interface across operating systems.
Visit Ollama →Amazon Bedrock
Managed access to models from several labs inside AWS, with enterprise identity and networking controls.
Visit Amazon Bedrock →Azure AI Foundry
Microsoft's catalogue and deployment surface for models inside Azure, with governance tooling attached.
Visit Azure AI Foundry →Fireworks AI
Fast hosted inference for open models, with fine-tuning and function calling on the same platform.
Visit Fireworks AI →Groq
Very low-latency hosted inference for open models, running on its own purpose-built hardware.
Visit Groq →OpenRouter
A single API in front of many providers and models, useful for comparison and for failover between them.
Visit OpenRouter →Replicate
Run open models through an API without managing infrastructure, including many community fine-tunes.
Visit Replicate →Together AI
Hosted open-model inference plus training and fine-tuning on dedicated GPU clusters.
Visit Together AI →llama.cpp
Efficient C/C++ inference that runs quantised models on ordinary hardware, including laptops and small devices.
Visit llama.cpp →vLLM
High-throughput open-source inference server, the common default for serving open weights at scale.
Visit vLLM →No matches. Try a different search.
Advertisement
Questions
Short, direct answers.
What does 'open weights' actually mean?
That the trained parameters can be downloaded and run on your own hardware. It does not necessarily mean open source in the traditional sense — training data and code are usually not released, and many licences restrict commercial use or scale. Read the specific licence rather than assuming, because they differ substantially between families.
When is a hosted API better than running weights yourself?
Almost always at the start. An API gives you the newest capability with no infrastructure and no idle cost. Self-hosting wins when you have steady high volume, a hard data-residency requirement, or a need to guarantee the model will not change under you — that last one is underrated, because hosted models are updated and your prompts were tuned against the old behaviour.
Why do benchmark scores disagree with real-world results?
Benchmarks measure narrow, often contaminated tasks, and leaderboards reward optimising for them. They are useful for excluding clearly unsuitable models and close to useless for choosing between good ones. Build a small evaluation from your own real inputs — a few dozen examples will tell you more than any public leaderboard.
What is a context window and how much do I need?
It is how much text a model can consider at once, prompt and reply together. Bigger is not automatically better: attention quality often degrades well before the stated limit, and long contexts cost more and run slower. Retrieval that sends only the relevant passages usually beats pasting an entire corpus into a large window.
Does this directory take payment for listings?
No. Listings are editorial and unpaid. Advertising on the site is labelled and has no bearing on what gets listed or where it appears.