Zyphra ZAYA1 is an open-source large language model family released by Zyphra in May 2026, featuring a mixture-of-experts (MoE) architecture designed for efficient inference and resource-constrained deployment. The model represents a significant development in open-model initiatives, combining sparse activation patterns with multimodal capabilities to achieve competitive performance across language and vision tasks.

The ZAYA1 family employs a mixture-of-experts design that reduces computational requirements during inference through selective expert activation. The flagship variant, ZAYA1-74B-Preview, consists of 74 billion total parameters with only 4 billion parameters active during inference, achieving approximately a 94% reduction in computational overhead compared to dense models of equivalent total parameter count ¹⁾.

This sparse activation approach addresses a critical challenge in large language model deployment: the computational cost of inference scales linearly with active parameters. By routing inputs to specialized expert subnetworks, MoE architectures maintain parameter efficiency while enabling knowledge distribution across distinct domains and capabilities ²⁾.

The multimodal variant, ZAYA1-VL-8B, extends the architecture to vision-language tasks with 8 billion total parameters and 700 million active parameters. This configuration enables processing of both text and image inputs while maintaining the efficiency characteristics of the base model, addressing the growing demand for unified multimodal processing pipelines in production environments.

The ZAYA1 models were trained on AMD hardware infrastructure, representing an important step toward diversifying AI model training across different accelerator ecosystems. AMD's ROCm platform provides an alternative to NVIDIA-dominant training pipelines, potentially reducing dependency on single-vendor hardware constraints and improving accessibility for organizations with existing AMD infrastructure investments.

All ZAYA1 variants are released under the Apache 2.0 license, permitting commercial use, modification, and distribution with minimal restrictions. This licensing approach facilitates integration into production systems, academic research, and commercial applications without the complex compliance requirements associated with more restrictive open-source licenses ³⁾.

The open release of ZAYA1 models generated positive community engagement, with validation of both the underlying architecture and the training methodology. Early evaluation across standard benchmarks and downstream tasks confirmed that the MoE approach successfully balances parameter efficiency with task performance, validating the sparse activation strategy for production deployment scenarios.

The community-driven validation process contributed to establishing ZAYA1 as a reliable baseline for open-model development, demonstrating that mixture-of-experts architectures can achieve practical efficiency gains without sacrificing task generalization across diverse domains ⁴⁾.

ZAYA1's efficiency characteristics enable deployment in resource-constrained environments including edge devices, cost-sensitive cloud infrastructure, and research settings with limited computational budgets. The multimodal variant supports applications requiring simultaneous language and vision understanding, such as document analysis, image captioning, visual question answering, and embodied AI systems that must process both textual context and visual inputs.

Organizations implementing ZAYA1 benefit from the reduced inference latency and energy consumption associated with sparse activation, critical factors in real-time applications and large-scale inference scenarios where computational costs directly impact operational expenses ⁵⁾.

• Zyphra
• ZAYA1-74B-Preview
• ZAYA1-8B
• Zyphra AI
• Qwen 3.6 27B