Global semiconductor company AMD is supporting the training of frontier-level AI models by innovative AI startup Zyphra.

AMD announced today that it has successfully trained ZAYA1, the first large-scale Mixture-of-Experts (MoE) foundation model, in collaboration with Zyfra.

This achievement was achieved based on the AMD Instinct MI300X GPU, AMD Pensando networking, and the ROCm open source software stack, and is evaluated as a case that demonstrates AMD's technological excellence in the field of large-scale AI model training.

The ZAYA1 Base model developed by Zaipla has demonstrated competitive performance compared to existing open models in various areas such as reasoning, mathematics, and coding.

In particular, it outperformed Llama 3 8B and OLMoE in several benchmarks, and demonstrated performance comparable to Qwen3 4B and Gemma3 12B.

ZAYA1, with only 760 million active parameters out of a total of 8.3 billion parameters, achieved high-performance, low-cost learning through an efficient structure.

This was made possible by the 192GB of high-bandwidth memory of the AMD Instinct MI300X GPU, which enabled complex tensor sharding.g) Supported stable and fast learning without procedures.

The AMD Instinct MI300X GPU delivers performance and scalability optimized for large-scale AI workloads.

Complexity is reduced and throughput is improved with 192GB of high-bandwidth memory, and model storage time is shortened by approximately 10 times with AMD-optimized distributed I/O, ensuring efficient learning stability.

“AMD’s leadership in accelerated computing is enabling innovators like Xifra to push the boundaries of AI,” said Emad Barsoum, vice president of AI and Engineering, AMD Artificial Intelligence Group. “This achievement is a significant example of the flexibility and performance that AMD Instinct GPUs and Pentium Networking deliver.”

“Efficiency is a core philosophy at Zyfra, and is reflected in every process, from model architecture design and learning algorithms to hardware selection that provides the best price-performance ratio,” said Krithik Puthalath, CEO of Zyfra. “ZAYA1 is very significant as it is the first example of large-scale learning implemented on an AMD platform.”

He continued, “This achievement demonstrates the power of our approach of co-designing model architectures with semiconductors and systems,” adding, “Zifra will further strengthen its collaboration with AMD and IBM to build the next generation of multimodal foundation models.”