InclusionAI publishes Ling 3.1 Flash weights for serving and further training

InclusionAI published the Ling 3.1 Flash checkpoint and a separate base release.

InclusionAI published the Ling-3.1-flash-1019 checkpoint and a separate base checkpoint on October 27. The model card describes the first as a post-trained update to Ling-3.1-flash. Both give teams downloadable weights for experimenting with the model beyond the original hosted release, and the post-trained checkpoint is listed under the MIT license. [Post-trained checkpoint](https://huggingface.co/inclusionAI/Ling-3.1-flash-1019) · [Base checkpoint](https://huggingface.co/inclusionAI/Ling-3.1-flash-base)

What the two checkpoints are for

The post-trained version is the starting point for teams evaluating the behavior described in the release. The base checkpoint gives researchers a different starting point for studying how their own training changes the model. A result measured on one checkpoint does not establish the same result for the other. Record the checkpoint identifier and serving configuration when comparing the two.

The model card describes a mixture-of-experts architecture with 560 billion parameters overall, 25 billion active per token and a context window of up to one million tokens. Selecting a fraction of experts per token lowers the arithmetic that a request needs relative to activating the entire model. The serving system still has to make the weights available as requests route to different experts. InclusionAI's documented SGLang setup uses eight H200 GPUs, an illustration of the difference between active parameters and the hardware involved in hosting the model. [Model card and serving example](https://huggingface.co/inclusionAI/Ling-3.1-flash-1019)

Interpreting the reported gains

InclusionAI attributes the changes to further post-training. It reports a rise from 62% to 68% on AutomationBench-AA and from 54% to 57% on SciCode. Those are test-specific observations. The release also cites two versions of the Artificial Analysis index, so a before-and-after comparison needs scores from the same version. [Published evaluation results](https://huggingface.co/inclusionAI/Ling-3.1-flash-1019)

A serving evaluation can start with the length and concurrency the intended application actually requires. Check whether the model retrieves the right information near the beginning and end of long inputs, then track response time and accelerator use as concurrent traffic grows. The downloadable checkpoints make those experiments possible on an operator's own infrastructure. They leave the cost and quality of a specific deployment to be measured there.