The Open Source AI Movement Is Winning. Here's the Data.

01 The Closing Gap

In January 2024, the best open-weight model (Llama 2 70B) scored 63% on MMLU. The best proprietary model (GPT-4) scored 85%. The gap was 22 points. In July 2025, the best open model (DeepSeek V3) scores 88%. The best proprietary model (GPT-4.5) scores 92%. The gap is now 4 points. At this rate of closure, open models will match proprietary frontier models by mid-2026.

02 The Meta Strategy

Meta's decision to open-source Llama was strategic, not charitable. By giving away the model, Meta commoditized the model layer — the thing OpenAI and Google were charging for. If models are free, value moves to data, compute, and application integration — areas where Meta (with 3 billion users) has structural advantages. It's the same strategy Google used with Android against Apple's iOS.

03 The DeepSeek Disruption

DeepSeek V3 was trained for approximately $5.6 million. GPT-4 cost an estimated $100 million+. The gap in training cost is closing even faster than the gap in quality. This means the competitive moat of having 'the best model' is shrinking. The companies that win will be those with the best data pipelines, the most efficient inference, and the deepest application integration — not necessarily the biggest model.