Databricks says DBRX, its latest open-source LLM, outperforms its open-source competitors, potentially offering enterprises a low-cost way to train it on their own data for generative AI use cases.
Data lakehouse provider Databricks has released a family of open-source large language models (LLM), DBRX, that it says outperforms OpenAI’s GPT 3.5 and open-source models such as Mixtral, Claude 3, Llama 2, and Grok-1 on standard benchmarking tests.
DBRX can be downloaded for free from GitHub and Hugging Face for research or commercial use.
This provides enterprises the opportunity to not only reduce their cost of developing generative AI use cases with their own enterprise data without being held back by constraints put forth by providers of closed models, such as OpenAI, on commercial usage.
The strategy to launch DBRX can be traced back to April last year, when the company launched its first open source LLM, Dolly 2.0, to showcase that enterprises had alternatives to models such as GPT 3.5 and GPT-4.
DBRX is supported on AWS, Google Cloud, and on Microsoft Azure via Azure Databricks, so enterprises can download the model and run it on graphical processing units (GPUs) wherever they wish.
Alternatively, enterprises can also choose to subscribe to DBRX and additional tools, such as retrieval augmented generation (RAG), for customizing the LLM via Databricks’ Mosaic AI Model Serving offering.
Mosaic AI Model Serving connects to DBRX via what the company calls Foundation Model APIs, which allows enterprises to access and query LLMs from a serving endpoint.
The Foundation Model APIs are provided in two pricing modes—pay per token and provisioned throughput.
While the pay per token is billed on the basis of concurrent requests, throughput is billed per GPU instance per hour. Both the rates, including cloud instance cost, start at $0.070 per Databricks unit.
The company also provides a pricing band for different GPU configurations.
As part of the LLM launch, Databricks has released two models under an open license with certain restrictions: DBRX Base, a pretrained base model, and DBRX Instruct, a fine-tuned version for few-turn interactions.
DBRX is also expected to be available through the Nvidia API Catalog and supported on the Nvidia NIM inference microservice.
While DBRX outperforms most models available today, according to Databricks’ tests, OpenAI’s GPT-4 leaves it behind on most benchmarks.
