Welcome, Argmax!
Foundation models on device.
- Founder: Atila Orhon
- Sector: AI / ML
- Location: San Francisco, CA
The Opportunity
At Salesforce Ventures, we’re optimistic that nearly every enterprise will have deployed AI in production by 2025. Further, we anticipate that as more use cases move to production, inference will dominate AI operating expenses instead of experimentation or training. In turn, the constraints of server-side inference will become more acute, which presents cost, latency, use case, and data privacy limitations.
At the same time, the commodity hardware made available to users via their own devices has never been more capable and performant. We’ve entered a world where user devices can, in many cases, support model inference but, until recently, haved lacked the necessary infrastructure.
“On-device LLMs are one of the next frontiers of AI. Shifting the deployment of LLMs directly on devices, customers can enjoy private, fast, and offline generative and agentic AI experiences.” – Shelby Heinecke (Salesforce AI Research)
The Solution
Argmax enables AI developers to seamlessly deploy and run state-of-the-art models and commercial-scale inference workloads directly on user devices. Beginning with open source, Argmax is tackling the blockers preventing on-device from becoming industry standard one by one. Their first release, WhisperKit, provides developers with free and virtually error-free translation and transcription to enable on-device speech recognition. Their second release, DiffusionKit, allows developers to run inference for Diffusion models on device. Both tools greatly simplify AI deployment for developers today and are already utilized by Salesforce and several other enterprises that are developing mobile AI use cases that require supporting on device infrastructure.
Over time, the Argmax platform will grow to support an ecosystem of large and small models deployed both on the server and on user devices, intelligently routing inference workloads to optimize for latency, privacy, and accuracy.
Why We’re Backing Argmax
The importance of on-device inference may not yet be obvious in a world where many AI use cases are still pre-production. However, at Salesforce, many of our users are mobile-first when in the field, and being able to serve valuable speech-to-text or real-time transcription features directly on device, wherever users are, is critical in the near-term. However, building the infrastructure to broadly support inference across user devices is no easy task, and few teams are equipped to do it.
Argmax was founded by CEO Atila Orhon alongside team members Brian Keene, and Zach Nagengast — all of whom greatly contributed to the efficient deployment of AI on device at Apple. This trio brought the developer ecosystem along with them through their open source contributions, including Apple’s Neural Engine Transformers and key architecture contributions to Apple’s Core ML private inference engine. Atila, Brian, and Zach possess an opinionated view on what it will take to make on-device inference the standard, and have assembled an incredible team to carry out their vision.
What’s ahead?
We’re excited to work with Atila and the Argmax team to push the boundaries of what’s possible at the edge and enable an AI experience that is not only lower cost and more performant, but also potentially more personalized and secure. It’s still early innings for AI, but we believe Argmax is incredibly well-positioned to serve as a key piece of infrastructure for the AI-forward enterprise.
We’re also excited for Argmax to be our first investment out of our newly minted $500M AI Fund.
Welcome to Salesforce Ventures, Argmax!