Ranked by benchmarks, popularity, efficiency, and versatility. Compare open-source models and find what runs on your hardware.
Benchmark data last synced Jun 6, 2026
| Runs on | Compare | ||||||
|---|---|---|---|---|---|---|---|
AA75.4 | Alibaba | 60 | 35B(3B active) | 262K | 8.5 GB | 12GB+ | |
AA71.7 | Alibaba | 53 | 35B(3B active) | 262K | 8.5 GB | 12GB+ | |
AA71.0 | Alibaba | 68 | 397B(17B active) | 262K | 46.0 GB | 48GB+ | |
AA70.2 | Moonshot AI | 70 | 1000B(32B active) | 262K | 86.2 GB | 96GB+ | |
AA70.1 | Alibaba | 63 | 27B | 262K | 72.8 GB | 80GB+ | |
BB68.1 | Google | 43 | 26B(4B active) | 256K | 11.0 GB | 12GB+ | |
BB67.6 | Moonshot AI | 65 | 1000B(32B active) | 256K | 84.6 GB | 96GB+ | |
BB66.8 | Alibaba | 63 | 27B | 262K | 72.8 GB | 80GB+ | |
BB66.3 | DeepSeek | 57 | 671B(37B active) | 128K | 59.8 GB | 64GB+ | |
BB66.2 | Alibaba | — | 397B(17B active) | 256K | 45.2 GB | 48GB+ | |
BB64.1 | Technology Innovation Institute | 0 | 40B | 2K | 24.4 GB | 32GB+ | |
BB63.7 | NVIDIA | — | 7B | 4K | 4.8 GB | 8GB+ | |
BB63.6 | DeepSeek | 69 | 1600B(49B active) | 1M | 420.9 GB | 424GB+ | |
BB63.2 | Alibaba | 60 | 122B(10B active) | 262K | 27.3 GB | 32GB+ | |
BB62.8 | MiniMax | 59 | 230B(10B active) | 205K | 22.7 GB | 24GB+ |
The right AI model is the smallest one that solves your task, runs on hardware you can afford, and ships under a license your business can live with.
There are thousands of public models. Most of the work in picking one is matching size to your hardware, capability to your task, and license to your business. Open weights (Llama, Qwen, Mistral, DeepSeek) cover almost every general-purpose need within 6 to 12 months of frontier closed models.
This directory ranks models by benchmark performance, popularity, efficiency, and versatility, and shows the VRAM each model needs at every quantization level. Use the Find My Model wizard for a quick recommendation or the filters to narrow by provider, architecture, parameters, or VRAM budget.
Model Providers
Every AI lab that ships models we track, ranked by model count, average score, and frontier presence. Compare OpenAI, Anthropic, Meta, Google, Alibaba, and more.
Model Families
Every model series we track, grouped by family. See every Llama, Qwen, GPT, Claude, Gemini, and beyond on a single page.
An open-source AI model is one where the weights, and usually the training code, are public. You can download it, run it on your own hardware, and modify it. Most of the models here are open-source. Closed models like GPT-5, Claude Opus, and Gemini are included for benchmark reference but cannot be downloaded.
Open weights means the trained weights are public but the training data and code may not be; this covers Llama, Qwen, Mistral, and most "open" releases. True open source means weights, code, and data are all public; OLMo and a few others qualify. Proprietary means closed, API-only access; OpenAI, Anthropic, and Google flagships fall here.
For general chat, the top Qwen, Llama, and DeepSeek text models are within reach of GPT-4 quality. For code, DeepSeek Coder, Qwen Coder, and CodeLlama dominate. For local inference on a laptop, look at 7B to 13B variants. The "Find My Model" wizard on this page asks three questions and points you at a short list.
Each model page shows the exact VRAM the model needs at FP16, 8-bit, and 4-bit quantization, plus a compatibility table for every GPU and Mac in our hardware directory. A 7B model fits in 8GB VRAM at 4-bit. A 70B model fits in 48GB. Anything larger needs multi-GPU or a high-memory Apple Silicon Mac.
Each model gets a composite score from four factors: benchmark performance (45 percent), popularity from downloads and likes (25 percent), efficiency in score per gigabyte of VRAM (20 percent), and versatility across capabilities (10 percent). Weights differ slightly per modality. Click the scoring methodology link in the table header to see the full formula.
Model metadata, downloads, and benchmark scores sync from Hugging Face, Ollama, and the public leaderboards on a daily cron. New launches usually appear within a day. The "Catalog refreshed" stamp above shows when the latest sync completed.
Closed-source flagships are listed for benchmark comparison but you cannot run them locally. Open the comparison panel to put one or two open-source models next to GPT-5, Claude, or Gemini and see the benchmark gap. The /models/api-prices page tracks live API pricing for closed models.
