Ant group unveils Ling-3.1-flash AI model with 560b parameters and 1m context support

On October 16, 2026, Ant Group announced the open-sourcing of its Ling-3.1-flash AI model, which boasts approximately 560B total parameters and 25B active parameters per token, with support for up to 1M context length. In public evaluations by Artificial Analysis, the model scored an intelligence index of 41 and achieved an output speed of about 210 tokens per second. For local deployment, the team applied MXFP4 quantization to routed experts and kept dense layers in block-FP8, reducing weight usage from 534GB to 292GB. The model runs efficiently on 4 Hopper GPUs instead of 8, offering better performance for teams with server-level GPUs. The model also features parallel inference solutions for long-context processing and enhanced JSON Schema support for tool calls, facilitating integration with existing business interfaces. In simulated production requests, the model achieved an average first token latency of 604 milliseconds and an average token interval of 3.28 milliseconds on an 8-card Hopper setup. For capacity planning, the model reached 1.36 QPS on a 4-card Hopper setup, meeting latency constraints for real-world deployment. Ant Group also offered discounted pricing for Ling-3.1-flash through OpenRouter, Ling Studio, and Hugging Face, with rates starting at $0.075 for input tokens and 0.5 yuan for input tokens in Chinese currency.