Ant Group Open-Sources Ling-3.1-flash, a 560B-Parameter Model with 1M Context

On October 15, Ant Group's Bailing announced the open-source release of Ling-3.1-flash, a Mixture-of-Experts model with approximately 560 billion total parameters and 25 billion activated per token, supporting up to 1 million tokens of context. In Artificial Analysis public benchmarks, the model scored 41 on the intelligence index and delivered an output speed of about 210 tokens per second.

For local deployment, the team applied MXFP4 quantization to routed experts and block-FP8 for dense layers, reducing the weight footprint from about 534GB to 292GB. The corresponding Hopper GPU configuration dropped from 8 to 4 cards, though actual requirements vary with context length, concurrency, and service setup.

For long-document processing, the team provides parallel context and parallel decoding (PCP/DCP) solutions, and enhanced tool-calling support for JSON Schema to integrate with existing business interfaces. In a simulated production workload with about 25.47K input tokens, about 224 output tokens, and a 76% cache hit rate, the 8-card Hopper setup achieved an average time-to-first-token of 604 ms and an average time-between-tokens of 3.28 ms.

In a 4-card Hopper throughput test with about 8.5K input tokens and about 800 output tokens, the model reached about 1.36 queries per second while meeting constraints of P50 time-to-first-token under 3 seconds and P90 inter-token latency under 70 ms. Ling-3.1-flash is currently offered with a limited-time 75% discount: $0.075 input, $0.225 output, and $0.015 cache read on OpenRouter; RMB 0.5 input, RMB 1.5 output, and RMB 0.1 cache read on Ling Studio.

The model is available on Hugging Face and ModelScope.