Ant Group Open-Sources Ling-3.1-flash Model with 1M Context
Ant Group has open-sourced the Ling-3.1-flash AI model, which features a 560B Mixture-of-Experts (MoE) architecture with 25B active parameters per token and a 1M context window. The model leverages advanced MXFP4 and block-FP8 quantization to reduce memory overhead and support efficient local deployment. By open-sourcing a large-scale MoE model optimized for local deployment and long-context processing, Ant Group lowers hardware barriers for enterprises building proprietary AI applications. This release provides competitive performance for tasks like enterprise knowledge management while enabling more cost-effective self-hosted infrastructure. Quantization successfully reduced model weights from approximately 534GB to 292GB, allowing deployment hardware requirements on Hopper GPUs to drop from 8 cards to 4 cards. Additionally, the team integrated operator fusion and DSpark speculative decoding to achieve low latency, reporting a 604ms time-to-first-token in simulated production tests.
## BACKGROUND
Large language models often rely on Mixture-of-Experts (MoE) architectures to scale up total parameters while keeping active compute low per token, making them faster and cheaper to run. Techniques like MXFP4 quantization compress neural network weights aggressively to save VRAM, while long-context parallel inference methods such as PCP and DCP distribute massive token sequences across multiple GPUs.