A plain-English deep dive into every major open-weight AI family — who builds them, how they compare, how to run them, and which one is right for you.
150+Open Models
8AI Families
50Top Intelligence Score
$0To Self-Host
52MMonthly Ollama Downloads
Section 01
What Is Open Source AI?
Think of "open source AI" like a recipe that anyone can read, copy, modify, and use — as opposed to proprietary AI (like ChatGPT) where the recipe is secret. Open AI models release their weights (the trained brain of the model) so anyone can download and run them. This matters because it means privacy, no monthly fees, and the ability to customize the AI for any purpose.
Open vs. Closed: What's the Difference?
A side-by-side look at the key practical differences between open-weight and proprietary AI models.
Intelligence Gap Closing Fast
The best open-weight models (blue) vs. best proprietary models (green) on the Artificial Analysis Intelligence Index (0–57). Gap was 20+ points in 2023. Now near-zero.
The Key Milestones: From Niche to Frontier
2023 — The First Wave
Meta released Llama 2 for free, sparking the modern open-source LLM movement. Mistral released their tiny-but-powerful 7B model, showing smaller models could punch above their weight. Open source was interesting but clearly behind proprietary models.
Llama 2Mistral 7BFalconGemma 1
2024 — The Ecosystem Matures
Llama 3 matched GPT-3.5 level quality. DeepSeek V2 and V3 stunned the world: Chinese open-source models rivaling GPT-4 at 1/10th the cost. Qwen 2.5 proved Alibaba was a serious player. Tools like Ollama made local deployment accessible to non-engineers.
Llama 3.1 405BDeepSeek V3Qwen 2.5Gemma 2Mistral Large
2025 — The Reasoning Revolution
DeepSeek R1 (January 2025) triggered a $1 trillion stock market drop — a free, open-source model matching OpenAI's o1 reasoning model. Llama 4 Scout arrived with the world's largest context window (10 million tokens). Qwen 3 introduced "thinking mode" to open-source. Chinese lab downloads overtook US models worldwide.
DeepSeek R1 ⭐Llama 4 ScoutQwen 3Kimi K2Mistral Small 3.1
2026 — The Gap Closes (Current)
"China Open Model Week" (Feb 2026): GLM-5, MiniMax M2.5, and Kimi K2.5 released within days — all MIT-licensed, all competitive with Claude Opus and GPT-5. GLM-5 hit the #1 open-source spot on the Artificial Analysis Intelligence Index with a score of 50. The performance gap between open and closed AI has effectively closed for most use cases.
GLM-5 ⭐MiniMax M2.5/M2.7Kimi K2.5Qwen 3.5Llama 4
Section 02
The Major Open Source AI Families
An AI "family" is a brand with multiple versions, like iPhone models. Each family has a maker, a specialty, and a license that determines how freely you can use it. Here's every major open-source AI family as of March 2026.
Section 03
Open Source Leaderboard — March 2026
The top open-weight models ranked by intelligence, including benchmark scores, license type, hardware needed to run locally, and what each model is best at. Models with an "S" tier compete directly with the best proprietary AI systems available.
#
Model
Family
Tier
Intelligence Score
SWE-Bench
AIME 2025
License
Min. VRAM
Best For
Section 04
Intelligence Trends by Family
How each open-source AI family's best model has improved year over year. The "Intelligence Score" is from Artificial Analysis — think of it like a GPA across 10 different subjects. Dashed lines are projected 2027 estimates based on observed growth rates.
All Open Source Families — Intelligence Over Time
Each line represents the peak score of that family's best model per year. Dashed = 2027 projection.
Best Open Source vs. Best Proprietary — Current Generation (2026)
Top models from each open-source family compared to the frontier proprietary models (gray bars). Score out of 57.
Intelligence Score vs. Cost ($/1M tokens) — Open Source Models
Top-right quadrant = best value. Open models clustered at low cost; some now reach proprietary quality levels.
Section 05
Run AI on Your Own Computer
You don't need a supercomputer. In 2026, a standard laptop can run a capable AI model completely offline — no internet, no subscription, total privacy. Here's what hardware you need at different budget levels, and which models work best on each.
What the Terms Mean (Plain English)
Glossary: The Words You'll See
No jargon. Here's what common open source AI terminology actually means.
Quantization: Quality vs. File Size Trade-Off
Quantization shrinks models so they fit on your hardware. Higher = better quality but needs more memory.
Section 06
Tools to Run Open Source AI
You don't need to be a programmer to run open source AI. These four tools handle the complexity for you. Each one fits a different style of user — from "I hate the command line" to "I want full server control."
Section 07
Which Open Source AI Is Right for You?
Straight answers. No jargon. The best open-weight model for 10 different types of people and tasks — all models you can download for free.
Section 08
Research Quality Checklist
A self-audit verifying what this report accomplishes, what data was gathered, and any limitations in the research.