From GPT to Llama: Deploy Any AI Model On-Premises with AIBOX

The AIBOX series of products have the characteristics of high performance, low power consumption, and strong environmental adaptability. The computing power covers 6-157 TOPS. By combining diversified deep learning algorithms, it is compact in size and can support the private deployment of mainstream large models, empowering digitalization for multiple smart industries.

Currently, Firefly has launched a total of 9 AIBOX products, which are adapted to application scenarios in different industries through different computing power, energy efficiency, appearance, etc.

Firefly x NVIDIA

Equipped with NVIDIA's original Jetson Orin series core module, it is compatible with accelerated computing functions of various performance and price points, with a computing power of up to 157 TOPS; it also supports the NVIDIA software ecosystem. With its powerful computing performance, excellent energy efficiency and easy development experience, it can meet the needs of various independent applications.

Large Language Model

• Robot Model: supports ROS robot model.
• Language Model: supports large models under Transformer architecture, such as Llama2/ChatGLM, Qwen and other large language models for private deployment.
• Visual Model: supports ViT, Grouding DINO, SAM and other large visual models for private deployment.
• AI Painting: supports the private deployment of Stable Diffusion V1.5 image generation model in the AIGC field.

Firefly x Rockchip

Equipped with Rockchip's flagship AIoT chip, it adopts a big-small core architecture with a maximum main frequency of up to 2.4 GHz, providing powerful hardware support for high-performance computing and multi-tasking. At the same time, this series has industrial features such as low power consumption and long battery life, which is adapted to the needs of industrial application scenarios.

Large Language Model

• Supports large models under the Transformer architecture, such as Gemma, Llama2, ChatGLM, Qwen, Phi and other large language models for private deployment.

Firefly x SOPHON

This series is equipped with SOPHON series AI processors, which are extremely cost-effective. The AIBOX-1684X has a computing power of up to 32 TOPS, supports mainstream programming frameworks, video encoding and decoding, and can be applied to artificial intelligence reasoning in cloud and edge computing applications.

Large Language Model

• Supports the private deployment of large models under the Transformer architecture, such as Llama2, ChatGLM, Qwen and other large language models.
• Supports the private deployment of large visual models such as ViT, Grouding DINO, SAM.
• Supports the private deployment of the Stable DiffusionV1.5 image generation model in the AIGC field.

Comprehensive AI Privatization Deployment

Most of the AIBOX series can support the private deployment of modern mainstream large models, such as the Gemma series, Llama series, ChatGLM series, Qwen series and other large language models.

• Supports traditional network architectures such as CNN, RNN, LSTM, etc.
• Supports multiple deep learning frameworks such as Caffe, TensorFlow, PyTorch, MXNet, and supports custom operator development.
• Supports Docker container management technology for easy image deployment.

Support Video Codec

The AIBOX series basically supports video encoding and decoding, and can support up to 8K@60fps video decoding and 8K@30fps video encoding. It supports simultaneous encoding and decoding, high resolution and multi-channel decoding capabilities, allowing large models to quickly obtain information in the video, providing richer data for model training and reasoning, improving visual analysis accuracy, and accelerating algorithm training and optimization.