Llama 3: Has Open Source AI Finally Caught Up to ChatGPT?
Meta released Llama 3 on April 18, 2024—8B and 70B models trained on 15 trillion tokens, rivaling GPT-4 class performance.
On April 18, 2024, Meta released Llama 3—and finally, open source AI matched closed models on major benchmarks.
The 70B model competed directly with GPT-4 class performance. The 8B model outperformed previous 13B+ models. Open source had caught up.
What Was New
Training data: 15 trillion tokens (7x more than Llama 2) Model sizes: 8B and 70B parameters initially Performance: Beats Gemini Pro, rivals GPT-4 on many tasks Multilingual: Better non-English performance 8K context: 4x longer than Llama 2
Plus: a 400B model was teased for later release.
The Benchmark Results
Llama 3 70B matched or beat:
- GPT-3.5 across the board
- Gemini Pro on most tasks
- Claude Sonnet on several benchmarks
For open source, this was unprecedented.
The Impact
Developers: Could now run GPT-4-class models locally Enterprises: No vendor lock-in, full control Startups: Build products without API costs Researchers: Advance AI without corporate gatekeepers
Open source was finally competitive on capability, not just cost.
Where Are They Now?
Llama 3.1 (July 2024) brought the 405B model, genuinely rivaling the best closed models. Llama became the foundation of the open-source AI ecosystem.
April 18, 2024 was when open source AI stopped being "good for open source" and became "genuinely competitive with anything."