One Model, Two Scores: 44 on the Old Test, 29 on the New

Ant Group's AI arm, InclusionAI, has published the downloadable weights of Ling-3.1-flash-1019 and its base model, Ling-3.1-flash-base, and both Hugging Face model cards label them MIT-licensed, one of the most permissive terms in software. ^[1][2]

The company says the updated model scored 44 on the Artificial Analysis Intelligence Index, up from 41 for its predecessor. ^[1] That number is real as a claim, but it belongs to an older version of the test. The evaluator's own page now lists the same model at 29. ^[3]

What was actually released

This is a release of weights, the learned numbers that let a model run, not a login to a hosted chatbot. Anyone can download them, run them on their own machines and adapt them.

There are two models. Ling-3.1-flash-1019 is described as an upgraded, post-trained version of Ling-3.1-flash, meaning it was trained further after its first build to shape how it responds.

Its card lists 560 billion total parameters, of which 25 billion are active for each token, the word fragment a model reads or writes at a time. It also lists a context window of up to one million tokens. ^[1] In plain terms, only a small slice of the model does the work on any given step, and it can take in very long documents.

The base model is a different animal. Its card calls it a foundation checkpoint.

A foundation model is the broad first-stage model that gets specialised later for particular jobs, and a checkpoint is a saved snapshot of a model's learned settings. This one is meant for continued training, adapting to a particular field and post-training research.

The card says it is not recommended as-is for direct end-user chat or for safety-critical use without further evaluation and adaptation. ^[2] It is raw material, not a product.

What the licence allows

The MIT licence is short and blunt. It lets anyone use, copy, modify, distribute, sublicense and sell the software, commercial use included. ^[4] A company can take the base model, train it on its own data, and sell the result without asking Ant for permission or paying it a fee.

There are two strings. Copies or substantial portions must carry the original copyright and licence notices. And the licence disclaims all warranties and liability, so if a model built on this checkpoint misbehaves, the licence gives the user no claim against the people who wrote it. ^[4]

That combination is why the licence matters more than the benchmark. Many model releases come with custom terms that limit certain uses or require approval at scale. MIT carries no such conditions beyond the notices. For the base model in particular, whose stated job is to be changed, the right to modify and redistribute is the whole point. ^[2]

Open, but not unlimited

Two limits are worth stating. The cards establish that the weights are available and give technical and benchmark information. They do not establish that the full training data or the entire training process is published. Open weights are not the same as an open recipe.

The second limit is hardware. The card's recommended serving instructions describe a setup with eight H200 GPUs, a class of data-centre accelerator. ^[1] That is not a laptop download. In practice the first people to build on this release will be companies, labs and universities with serious compute, or those who rent it.

The score and its asterisk

InclusionAI reports 44 on Artificial Analysis Intelligence Index v4.3.2, against 41 for the previous Ling-3.1-flash. The same card reports a rise on AA-Briefcase from about 1,400 to 1,524 Elo, a rating system borrowed from chess in which models are ranked by head-to-head results, and on AutomationBench-AA from 62% to 68%. ^[1] All of these are the developer's own statements.

Artificial Analysis, which runs the index, now describes version 5.0, built from nine evaluations, and lists Ling 3.1 Flash 1019 at 29. ^[3] The two numbers cannot be ranked against each other, because they come from different versions of the test with different components. The 29 does not mean the model got worse, and the 44 is not a verdict that holds across versions.

What the index is meant to capture is clearer. Artificial Analysis describes it as a composite covering knowledge work, automated workflows, coding, scientific tasks and reasoning, measured independently by the company. ^[3] That independence is the reason a developer wants its number on the card. It is also the reason readers should ask which version of the number they are looking at.

It is not yet known when the two repositories first went public, whether outsiders can reproduce the 44, or how the model compares with others under the current index using the same settings. The cost of running very long inputs near the one-million-token limit is also unanswered.

The next real evidence will not come from a model card. It will come from what other people do with the base checkpoint once they start changing it.