The ecosystem of open-source (and open-weights) Large Language Models is developing rapidly. For developers, these models offer an independent, locally runnable, and cost-effective alternative to closed APIs. Below you will find an overview of the most influential models and their specific characteristics.
Llama, developed by Meta, is the undisputed pioneer among powerful open-weights models. The third generation (Llama 3) offers models ranging from 8B to 70B (with 400B in development) parameters.
The European startup Mistral AI surprised the market with highly efficient models. Their 'Mixtral' models utilize a Sparse Mixture of Experts (MoE) architecture, which ensures high performance with relatively low computing power requirements during inference.
The Qwen model family performs remarkably well across a wide range of sizes (from 0.5B to 72B). Qwen 2 has made significant strides in mathematical capabilities and multilingualism.
Gemma is built on the same technology and research as Google's Gemini models. Gemma 2 (available in sizes such as 9B and 27B) excels in its weight class and offers robust performance on standard consumer hardware.
Want to see how these open-source models perform against each other in specific tasks? Check out our AI Benchmark Hub for objective comparisons and the latest evaluations of the LLMnet ecosystem.