InclusionAI publishes Ling-3.1-flash weights under the MIT license

InclusionAI has published Ling-3.1-flash on Hugging Face, giving developers downloadable text-generation weights under the MIT license alongside a hosted inference option. The release offers two ways to access the model, but they are not equivalent commitments: the downloadable files total roughly 574 GB, while Hugging Face lists Novita as a live provider.

The download and the hosted route

The repository is public and ungated, with model weights supplied in Safetensors format. Its file metadata lists 573,801,615,144 bytes across the checkpoint. That is the size of the published model files—not a stated GPU-memory requirement or a complete serving configuration.

The checkpoint is tagged FP8, and its tensor metadata shows that most stored parameters use that format. It also includes tensors in BF16 and F32 formats. The download is therefore a mixed-format package, rather than a set of weights stored uniformly in one numerical format.

Hosted access is listed separately. Hugging Face marks Novita’s conversational inference service as live, with tool calling enabled. Its provider metadata marks structured output as unsupported. Those are the capabilities listed for that hosted integration; they should not be treated as a complete description of every possible deployment of the downloadable model.

Thinking controls and tool instructions

The supplied chat template includes an explicit thinking option. It accepts an enable_thinking setting, translates that into an on-or-off instruction, and defaults to thinking on when no alternative is supplied. The template also handles reasoning content separately from the final answer and preserves that content in the conversation formatting.

Tool use has its own instructions and message structure. When functions are provided, the template passes their signatures to the model and specifies how to format function names and arguments. It tells the model to point out missing parameters or cases where none of the supplied functions can be used. These are instructions in the template, not evidence that the model follows them reliably.

Read the scores with their harness

The model page publishes several evaluation results, all marked unverified in Hugging Face’s metadata. Among the listed scores are:

• SWE-bench Pro: 68.74, with a note that the run used the LingCode harness.
• Terminal Bench 2.1: 81.18.
• Deep SWE: 59.7, also marked as a LingCode harness run.

InclusionAI describes LingCode as a harness built on the open-source coding agent Pi. It is designed for coding jobs that run for hours, involve hundreds of tool-call rounds and repeatedly compact their context. The company says the harness improves some coding-agent results and is considering releasing it as open source.

That distinction matters when reading the coding scores: the page identifies some results with a particular agent setup, not just the model name. The weights are publicly available now; InclusionAI’s description of potentially open-sourcing LingCode remains a future possibility, rather than a commitment or an accompanying harness release.