2023
Llama, Mistral, and open source AI
Meta launches Llama and later Llama 2 as open source LLMs; Mistral AI from Paris releases Mistral 7B that outperforms closed models — open source AI comes of age.
Open source strikes back
While OpenAI, Anthropic, and Google competed with closed frontier models, 2023 was also the year open source AI came into its own. Meta released Llama in February 2023 — a family of language models from 7 to 65 billion parameters, initially intended for researchers. The weights were leaked within days and quickly spread across the internet. The open-source AI community had its first competitive foundation model.
Llama 2 and Meta's strategy
In July 2023, Meta released Llama 2 officially as open source, with commercial use permitted. This was a strategic decision: Meta had no direct commercial interest in selling AI services, but a strong interest in a healthy open-source AI ecosystem that prevented lock-in to competitors' proprietary models. Llama 2 was competitive with GPT-3.5 on many benchmarks and became the most widely used open-source LLM.
Mistral: small but mighty
In September 2023, Mistral AI — a Paris-based startup founded by former DeepMind and Meta researchers — released Mistral 7B. Despite having only 7 billion parameters, it outperformed Llama 2 13B on most benchmarks, demonstrating that architectural improvements and efficient training could compensate for model size. Mistral's open release approach and European roots made it a symbol of the possibility of competitive AI outside the American tech giants.
Sources
- Touvron, H. et al. (2023). Llama 2: Open Foundation and Fine-Tuned Chat Models. arXiv:2307.09288.
- Wikipedia — Llama
- Wikipedia — Mistral AI