Databricks

On March 27, 2024 Databricks released DBRX, a 132-billion-parameter mixture-of-experts large language model, under a permissive open-weight license including commercial use rights. At release DBRX outperformed Llama 2 70B and Mixtral 8x7B on standard benchmarks and was among the largest fully open-weight models of its time. Databricks published the training methodology, evaluation data and tokenizer alongside the weights, contrasting with closed releases from OpenAI, Anthropic and Google.

On April 12, 2023 Databricks released Dolly 2.0, the first instruction-following large language model with a license permitting commercial use, trained on a 15,000-prompt instruction dataset (databricks-dolly-15k) that Databricks released to the public domain under CC-BY-SA. The release accelerated commercial-grade open LLM availability and established a template for openly-released training datasets at a time when most commercial-grade instruction tuning relied on OpenAI API outputs (a use forbidden by OpenAI's terms).