Ant Group's Bailin announced the open-source release of the Ling-3.1-flash model yesterday. Ling-3.1-flash has approximately 560 billion total parameters, activates about 25 billion parameters per token, and supports a maximum context length of 1 million.
In the public evaluation by Artificial Analysis, Ling-3.1-flash has an intelligence score of 41 and an output speed of approximately 210 tokens per second.
For on-premises deployment, the team applied MXFP4 quantization to routed experts while keeping dense layers in block-FP8, reducing weight storage from approximately 534 GB to 292 GB. On the corresponding Hopper GPU test configuration, service deployment was reduced from 8 GPUs to 4 GPUs, enabling teams with server-grade GPUs to run the model with fewer cards; actual requirements still vary depending on context length, concurrency, and service configuration.
For long-context processing, the team provides a parallel context processing solution (PCP/DCP); tool calling has also been enhanced with improved support for JSON Schema, facilitating seamless integration with existing business interfaces.
Enterprise knowledge Q&A and office assistants often need to process lengthy documents, yet users still expect rapid responses. The team reduced inference overhead through operator fusion and DSpark speculative decoding: in simulated production requests with an average input of 25.47K tokens and output of 224 tokens, and a cache hit rate of 76%, the average time to first token on an 8-Hopper-GPU configuration was 604 milliseconds, with an average interval of 3.28 milliseconds between consecutive output tokens. This provides a benchmark for interactive response after long-document input.
Business systems that require continuous service must also evaluate how many requests a single instance can handle within a fixed GPU budget. In a 4-GPU Hopper throughput test with an average input of approximately 8.5K tokens and output of 800 tokens, the model achieved approximately 1.36 QPS while meeting the test constraints of a P50 first-token latency of less than 3 seconds and a P90 inter-token interval of less than 70 milliseconds. This provides a reference for capacity planning in self-deployed environments.
For the online experience, Ling-3.1-flash is currently offering a limited-time 2.5x discount, with the discounted prices as follows:
- OpenRouter (USD-denominated): Input $0.075; Output $0.225; Cache read $0.015.
- Ling Studio (CNY pricing): Input 0.5 CNY; Output 1.5 CNY; Cache read 0.1 CNY.
IT之家 provides an open-source link:
- Hugging Face: https://huggingface.co/inclusionAI/Ling-3.1-flash
- ModelScope: https://modelscope.cn/models/inclusionAI/Ling-3.1-flash
Related reading: "Antelope Releases Ling-3.1-flash Model: Approximately 560B Total Parameters and 25B Activated Parameters"