SambaNova Systems intros Samba-1 generative AI model
The new model is aimed at enterprises wanting to use generative models without compromising privacy. It also uses the best open source model for a given application.
AI hardware and software provider SambaNova Systems on Wednesday introduced a new 1 trillion-parameter generative AI model, Samba-1.
Samba-1 enables enterprises to train private data in a private environment, whether on the cloud or on premises. The model uses a variety of open source models, including Meta Llama 2, Mistral 7B and Falcon 40B.
Samba-1's technology takes these models and pulls the best models to answer a user's prompt based on the enterprise application, according to the independent vendor. For example, if a user asks the model to generate a nondisclosure agreement, Samba-1 chooses which of the open source models it is built on is best at creating legal NDAs and uses that model to perform the task.
The generative AI model is based on SambaNova's SN40L chip, which the AI vendor said is comparable to Nvidia AI chips from the AI hardware and software giant.
Samba-1 is now available within the SambaNova Suite, a full-stack generative AI platform optimized for enterprise and government organizations.
Growth of open source
SambaNova's new model comes as AI users increasingly welcome open source models.
For the most part, I think we're going to see more and more use of [open] models in the generative AI space.
Matthew EastwoodAnalyst, IDC
Most recently, cloud giants such as Microsoft and Google have also shown support for open source models. Over the last two weeks, Google introduced its open model Gemma, and Microsoft partnered with open source AI startup Mistral AI.
"For the most part, I think we're going to see more and more use of [open] models in the generative AI space," IDC analyst Matthew Eastwood said.
Therefore, SambaNova's use of open source models to create its generative model is not unusual, he said. However, open source models pose challenges such as lack of support from the vendors that create them.
"It's all the same challenge for folks that have to think about how they want to put guardrails in place to really make sure that these tools are being used in a way that's not going to cause them some harm downstream," Eastwood added.
Private vs. public
Samba-1 addresses a question enterprises face when deciding which generative AI approach they should take: open or not.
Many organizations want to use their own data to train AI models, largely for data privacy reasons. Those organizations tend to favor the more private models that Samba-1 provides.
The new Samba-1 model is built on the SambaNova AI chip and uses a compilation of different open source models.
"That doesn't mean that everything [should] be private and dedicated, but certainly for the high-strategic-value differentiated implementations that organizations are looking at ... SambaNova's basically providing an approach to delivering that," Eastwood said.
Samba-1 does this by enabling organizations with sensitive data to train models on that data within their private environment.
"This is as private as it gets," SambaNova co-founder Rodrigo Liang said. SambaNova provides an organization with a full-stack approach that brings the model right to where the data is so that training and inferencing are done in the same network, he said.
One challenge for SambaNova is ensuring it can support implementations of the open source models Samba-1 uses, Eastwood said.
"It could be models from the same organizations that they're leveraging here, but recognize that they're going to see lots of different potential model implementations come at them," he said. "They're going to need to have a good story for how they support those on their platforms as well."
Esther Ajao is a TechTarget Editorial news writer and podcast host covering artificial intelligence software and systems.
