Covering Disruptive Technology Powering Business in The Digital Age

image
DBRX, New Standard for Efficient Open Source Models, Officially Launched by Databricks
image

 

Databricks, the Data and Artificial Intelligence (AI) company, has announced the launch of DBRX, a general purpose large language model (LLM) that outperforms all established open source models on standard benchmarks.

This new offering democratises the training and tuning of custom, high-performing LLMs for enterprises so they no longer need to rely on a small handful of closed models. Available now, DBRX enables organisations around the world to cost-effectively build, train, and serve their own custom LLMs.

“At Databricks, our vision has always been to democratise data and AI. We’re doing that by delivering data intelligence to every enterprise—helping them understand and use their private data to build their own AI systems. DBRX is the result of that aim,” said Ali Ghodsi, Co-Founder of and CEO at Databricks.

Ghodsi added: “We’re excited about DBRX for three key reasons: First, it beats open source models on state-of-the-art industry benchmarks. Second, it beats GPT-3.5 on most benchmarks, which should accelerate the trend we’re seeing across our customer base as organisations replace proprietary models with open source models. Finally, DBRX uses a mixture-of-experts architecture, making the model extremely fast in terms of tokens per second, as well as being cost effective to serve. All in all, DBRX is setting a new standard for open source LLMs. It gives enterprises a platform to build customized reasoning capabilities based on their own data.”

DBRX Surpasses Open Source Models Across Industry Benchmarks

Databricks’s latest LLM outperforms existing open source LLMs like Llama 2 70B and Mixtral-8x7B on standard industry benchmarks, such as language understanding, programming, math, and logic.

DBRX Figure 1

Fig. 1. DBRX outperforms established open source models on language understanding (MMLU), Programming (HumanEval), and Math (GSM8K).

It also outperforms GPT-3.5 on relevant benchmarks.

DBRX Figure 2

Fig. 2. DBRX outperforms GPT 3.5 across language understanding (MMLU), Programming (HumanEval), and Math (GSM8K).

For an in-depth look at model evaluations and performance benchmarks, and to see how DBRX is competitive with GPT-4 quality for internal use cases such as SQL, visit the Mosaic Research blog.

Setting a New Standard for Efficient Open Source LLMs

DBRX was developed by Mosaic AI and trained on NVIDIA DGX Cloud. Databricks optimised DBRX for efficiency with a mixture-of-experts (MoE) architecture, built on the MegaBlocks open source project. The resulting model has leading performance and is up to twice as compute-efficient as other available leading LLMs.

It set a new standard for open source models, enabling customisable and transparent generative AI for all enterprises. A recent survey from Andreessen Horowitz found that nearly 60 percent of AI leaders are interested in increasing open source usage or switching when fine-tuned open source models roughly match performance of closed source models.

In 2024 and beyond, enterprises expect a significant shift of usage from closed towards open source. Databricks believes DBRX will accelerate this trend.

Paired with Databricks Mosaic AI’s unified tooling, DBRX helps customers rapidly build and deploy production-quality generative AI applications that are safe, accurate, and governed without giving up control of their data and intellectual property. Customers benefit from built-in data management, governance, lineage, and monitoring capabilities on the Databricks Data Intelligence Platform.

Availability

DBRX is freely available on GitHub and Hugging Face for research and commercial use. It is also available on AWS and Google Cloud, as well as directly on Microsoft Azure through Azure Databricks. It is also expected to be available through the NVIDIA API Catalog and supported on the NVIDIA NIM inference microservice.

To learn more, visit the Mosaic AI research blog or join the webinar on the 25th of April at 8:00 a.m. PT.

(0)(0)

Archive