Intel is optimizing SGLang for its chips with RadixArk

Intel has started a collaboration with RadixArk to optimize SGLang, an open-source inference stack, for Intel platforms. The November 12 release, issued at 3:30 p.m. Eastern from Santa Clara, frames the work as a way to deploy models on those chips with less custom engineering.

The release describes production inference as routing each prompt to the right compute. It calls an open runtime the link that makes a mixed stack practical, and the improvements go into community software rather than a separate deployment product.

The optimizations are going into the open SGLang framework. RadixArk chief executive Ying Sheng said the changes are being upstreamed into the core project, so a team does not have to choose between the open stack and a separate Intel build.

Intel is aiming the work at CPUs, GPUs, and accelerators. Anil Nanduri, a vice president in Intel's data center group, said developers should get Day-0 readiness on that hardware, including Intel Advanced Matrix Extensions. The release frames that as a choice of tools, not a requirement to use an Intel-only runtime.

Intel says the point is a shorter path to production and lower latency. It also points to deployment on data center and cloud capacity a customer already has. Those are the aims in the release. They are not a measured result.

The November 12 release does not include a speed figure. It does not name a model, a token rate, or a date when the optimizations land in a public SGLang build. The partnership is the announcement. The benchmark is not in it.