IT之家 IT之家 reported on October 16 that Ant Group's Bailian announced yesterday the open-sourcing of the Ling-3.1-flash model. Ling-3.1-flash has approximately 560B total parameters, activates about 25B per token, and supports up to 1M context.
In Artificial Analysis's public evaluation, Ling-3.1-flash scored an intelligence index of 41 with an output speed of about 210 tokens per second.
For local deployment, the team applied MXFP4 quantization to the routed experts while keeping the dense layers in block-FP8, reducing weight footprint from about 534GB to 292GB. In the corresponding Hopper GPU test configuration, service deployment was reduced from 8 GPUs to 4 GPUs, allowing teams with server-grade GPUs to run the model with fewer cards; actual requirements still vary with context, concurrency, and service configuration.
For long document processing, the team provides a long-context parallel inference solution (PCP / DCP); tool calling has also enhanced support for JSON Schema, making it easier to integrate with existing business interfaces.
Enterprise knowledge Q&A and office assistants often need to read lengthy documents first, while users still want replies as quickly as possible. The team reduced inference overhead through operator fusion and DSpark speculative decoding: in simulated production requests with average input of about 25.47K tokens, output of about 224 tokens, and a cache hit rate of 76%,8-GPU Hopper configuration had an average time to first token of 604 milliseconds, with an average interval between adjacent output tokens of 3.28 milliseconds. This provides a test reference for interactive responses after long document input.
Business systems that provide continuous service also need to evaluate how many requests a single instance can handle within a fixed GPU budget. With an average input of approximately 8.5K tokens and an output of approximately 800 tokens, the 4-GPU Hopper throughput test with average input of about 8.5K tokens and output of about 800 tokens, the model reached approximately 1.36 QPS, while meeting the test constraints of time to first token P50 under 3 seconds and interval between adjacent output tokens P90 under 70 milliseconds. This provides a reference for capacity planning in self-hosted deployments.
For online experience, Ling-3.1-flash currently offers a limited-time 2.5-fold discount (i.e., 75% off), with post-discount prices as follows:
OpenRouter(priced in USD): input $0.075; output $0.225; cache read $0.015.
Ling Studio(priced in RMB): input 0.5 yuan; output 1.5 yuan; cache read 0.1 yuan.
IT之家 hereby attaches the open-source links:
Hugging Face:https://huggingface.co/inclusionAI/Ling-3.1-flash
ModelScope:https://modelscope.cn/models/inclusionAI/Ling-3.1-flash
Related reading:
Advertising disclaimer: The external jump links contained in this article (including but not limited to hyperlinks, QR codes, passcodes, and other forms) are used to convey more information and save screening time; the results are for reference only. All IT之家 articles containing external links include this disclaimer.