San Francisco, October 15th, 2026 —Inferact, the company founded by the creators of vLLM to build the production inference layer for AI, today announced an engineering collaboration with AMD to further strengthen support for production and agentic inference on AMD Instinct™ GPUs in vLLM, the most widely adopted open-source inference engine, and in the Inferact Platform built on it.
Why this matters
vLLM's job is to run well on the best hardware available for the workload. This collaboration builds on AMD Instinct as a platform for production inference and aims to improve how efficiently vLLM runs on these accelerators. That work is becoming more important as inference traffic changes shape. New model architectures and workloads, from coding agents to long-running assistants, place growing demands on serving systems. Meeting those demands takes optimization across the stack, from kernels and scheduling to cache management.
What each company contributes
AMD is contributing access to the AMD Instinct platform, along with deep hardware and kernel expertise, AMD ROCm™ software engineering, and the AITER kernel library. Inferact and the vLLM community are contributing the engine: model integration, serving features, and day-0 enablement for new releases. The work is carried by the same engineers who run vLLM day to day and who operate the production inference traffic on Inferact's own platform, ensuring that the improvements are grounded in real production traffic. The two sides work from shared model-by-model roadmaps covering kernels, serving features, and model coverage. Everything the partnership produces will land in open source.
Already live
The collaboration has shipped with major open-model releases this summer. Beyond model launches, Inferact operates production inference on AMD Instinct clusters for customers of its BYOC platform. This joint collaboration and enablement ensures that vLLM is optimized for real, production agentic workloads.
The technical work
The partnership spans the vLLM serving stack on AMD Instinct platform, from low-level optimization through the serving features that production traffic depends on. The current focus is agentic production workloads: long-running sessions, long pre-fixes, and tool calling with validation. Every configuration will ship as a public, per-model serving recipe, so any team can start from a known-good deployment.
Each new model bring-up creates foundations upon which subsequent model releases can build.
What to expect next
The companies are extending day-0 coverage for new models, deepening kernel co-development for agentic workloads, and publishing per-model performance data as it lands. The standing goal is unchanged: when a frontier open model ships, it should run well on AMD Instinct the same day, in vLLM, with optimized performance on real production workloads.
About Inferact
Inferact is building the production inference infrastructure for AI, enabling companies to serve optimized inference on their own infrastructure. Founded by the creators and core maintainers of vLLM, the most widely adopted open-source LLM inference engine, Inferact makes it faster, cheaper, and easier to serve frontier AI models at scale. vLLM supports more than 300 model architectures and has more than 3,000 contributors, and is used by frontier labs, hyperscalers, and startups around the world. Learn more at inferact.ai.
Media contact: contact@inferact.ai
