#Japan #Tokyo #FlashLabs #OrcaRouter #GLM-5.3

FlashLabs Unveils OrcaRouter for GLM-5.3 on Mac

In a groundbreaking announcement, FlashLabs Inc., headquartered in Chiyoda, Tokyo, has introduced the OrcaRouter, an AI inference gateway specifically designed for the Japanese market. This innovative platform is capable of running the massive LLM (Large Language Model) known as GLM-5.3, developed by the American research organization, Continuum AI. What sets the OrcaRouter apart is its ability to execute the full version of GLM-5.3—boasting an impressive 743 billion parameters—on a single Mac, optimized for Apple Silicon.

Key Features of GLM-5.3-MLX

The newly launched GLM-5.3-MLX version is not a distilled version but the full model, tailored specifically for the infrastructural capabilities of Apple’s M3 Ultra chip. Noteworthy advancements include:

• - The implementation of a proprietary quantization method called OrcaSAQ, enabling a significant reduction in model size from around 1.51TB to approximately 322GB, achieving nearly 79% size reduction while maintaining about 80.2% accuracy compared to the full precision.
• - Four variants of the model are available—2-bit, 3-bit, 4-bit, and 6-bit—giving developers flexibility depending on memory capabilities. For instance, a Mac with 512GB of memory can smoothly run 2-bit and 3-bit variants, while a 4-bit can be executed in a fitting configuration.
• - The top-tier 6-bit version utilizes a distributed setup requiring two 512GB Macs or can run effectively on eight NVIDIA H200 units. This model has demonstrated a cosine similarity exceeding 0.9997 when compared to the FP8 version.

Context and Challenges

The demand for local execution of large language models is on the rise, primarily due to data privacy concerns and the need for reduced latency in real-world applications. However, many of the leading open weight models typically demand hundreds of gigabytes of memory, posing a significant barrier for everyday developers working with standard devices.

To address these challenges, OrcaRouter employs advanced quantization techniques aimed at compressing models without compromising performance. The OrcaSAQ method avoids the necessity of calibration data while ensuring high precision for critical tensors according to the model structure, striking a balance between performance and memory usage efficiently.

Previously in August 2026, FlashLabs released a version of the GLM-5.3-Flash (with 320 billion parameters) optimized for Apple Silicon as a step towards the current full model deployment. The full GLM-5.3-MLX now provides a comprehensive solution for developers seeking high-performing models without the hefty memory requirements.

About GLM-5.3

Z.ai’s GLM-5.3 serves as a flagship model with a unique mixture of experts (MoE) architecture, exhibiting approximately 743 billion total parameters and 40 billion active parameters, equipped for a context length of 1 million tokens. The variations offered by GLM-5.3-MLX give developers room to optimize based on application needs and execution environments, affirming FlashLabs’ commitment to innovative AI solutions.

Future Prospects

Looking ahead, FlashLabs aims to continue researching and developing quantization technologies that make large models more accessible on everyday devices. Moreover, there are plans to expand the model catalog further, enhancing the capabilities of the OrcaRouter.

About FlashLabs Inc.

Founded in July 2023, FlashLabs focuses on AI application research, emphasizing the automation and autonomy of sales and customer experience through Human-AI Hybrid technologies. The company's headquarters is located at 10-8 Ichibancho, Chiyoda, Tokyo. Their official website can be accessed at FlashLabs.

About Continuum AI

Continuum AI is a U.S.-based research organization dedicated to developing foundational infrastructures for the next decade of intelligent systems and is responsible for the development of OrcaRouter. Their website is Continuum AI.

Contact Information

For inquiries regarding this matter, please contact:
Koki Kobayashi, Marketing Department, FlashLabs Inc.
E-mail: [email protected]

This innovation contributes to the evolving landscape of AI applications, making state-of-the-art models readily available for local execution and enabling developers to harness their full capabilities.