AMD's Instinct MI355X accelerators are now running Kimi-K3, a massive 2.8 trillion parameter AI model. The deployment includes Day 0 support for advanced inference capabilities, meaning the hardware and software stack were ready for the model from the moment it launched.
The Model: Kimi-K3
Kimi-K3 is a large language model with 2.8 trillion parameters, placing it among the largest publicly known AI models. Its size suggests it can handle complex reasoning, code generation, and multilingual tasks, though the developers have not released detailed benchmarks. The model's deployment on AMD hardware marks a shift in the AI infrastructure landscape, where Nvidia's GPUs have traditionally dominated large-scale training and inference.
AMD's Day 0 Support
AMD says it provided Day 0 support for Kimi-K3 on the MI355X GPUs. That means the company optimized its ROCm software stack and drivers to run the model efficiently from the first day of availability. Day 0 support is a competitive differentiator for AMD, as it reduces the time developers spend porting models and tuning performance. For Kimi-K3, this likely includes support for mixed-precision inference, memory optimization, and kernel-level adjustments specific to the model's architecture.
The deployment signals that AMD's MI355X is being taken seriously for frontier-scale AI workloads. With 2.8 trillion parameters, Kimi-K3 requires massive memory bandwidth and compute capacity—exactly what the MI355X is designed to deliver. The move could pressure Nvidia, which has dominated the market for large model inference. It also suggests that the team behind Kimi-K3 is prioritizing hardware diversity and cost efficiency, as AMD's GPUs often offer competitive pricing and availability compared to Nvidia's offerings.
No further details about the Kimi-K3 model's training data, architecture, or specific performance metrics have been released. The developers have not announced a public API or open-source release. For now, the focus is on the hardware partnership and the ability to run such a large model on AMD silicon.
