Agnes 3.0 Flash

Agnes 3.0 Flash is a multimodal language model developed by Sapiens AI, an artificial intelligence research laboratory based in Singapore. It is positioned as a high-speed, low-latency model optimized for real-world agentic workflows, including task planning, multi-step tool orchestration, and autonomous coding. The model architecture supports processing for text, image, and video inputs, focusing on instruction following and consistent output generation.

The model utilizes a hybrid-attention architecture consisting of 72 decoder layers. This design employs a 3:1 ratio of recurrent "gated delta rule" layers to standard global attention layers. By utilizing recurrent mechanisms for the majority of its layers, Agnes 3.0 Flash reduces the memory overhead of the KV cache at long context lengths, maintaining performance across extensive input sequences.

Agnes 3.0 Flash is provided in two distinct configurations: a production-tier API version and an open-weight "Preview" checkpoint. The flagship production model features a 1,000,000-token context window and is built on a Mixture of Experts (MoE) architecture with approximately 320 billion total parameters and 18 billion active parameters. The open-weight Agnes-3.0-Flash Preview is a 33-billion parameter dense model released under the Apache 2.0 license, supporting a 262,144-token context window.

Capabilities of the model include support for adjustable reasoning levels, which allows for the configuration of reasoning effort (high, medium, low, or off) based on task complexity. The model is engineered for grounded execution and multi-modal understanding, supporting the development of AI agents and tool-driven applications.