Ling-3.1-flash-1019 Reports Gains in Reasoning and Agent Benchmarks

TL;DR

• inclusionAI’s Ling-3.1-flash-1019 is an upgrade to Ling-3.1-flash, with 560B total parameters, 25B active per token and a 1M-token context window.
• The repository reports AA-Briefcase rising from about 1400 to 1524 Elo and AutomationBench-AA climbing from 62% to 68%.
• CritPt moved from 18% to 26%, SciCode from 54% to 57%, and the Artificial Analysis Intelligence Index v4.3.2 from 41 to 44.
• Every figure comes from the model’s own README, so treat them as claims until someone else reruns them.

Ling-3.1-flash-1019 Posts Higher Scores Across Agent and Science Tests

InclusionAI has put out Ling-3.1-flash-1019, an updated version of Ling-3.1-flash. The model repository on Hugging Face describes it as a hybrid-linear mixture-of-experts model with 560B total parameters, 25B active parameters per token, and a 1M-token context window.

The README credits the jump to post-training. Its own words: “With further post-training optimizations, Ling-3.1-flash-1019 demonstrates significant improvements in reasoning and real-world agentic capabilities.”

Here are the reported before-and-after numbers. AA-Briefcase goes from approximately 1400 to 1524 Elo. AutomationBench-AA moves from 62% to 68%. CritPt climbs from 18% to 26%, and SciCode from 54% to 57%.

On the Artificial Analysis Intelligence Index v4.3.2, the overall score rose to 44 from 41. That’s the single-number summary, and it’s a modest one.

Small Index Move, Bigger Story Underneath

Three points on an index doesn’t sound like much. But look at where the movement landed. The agent-style tests and the hard science test shifted the most, while SciCode crept up only three points.

CritPt is the eye-catcher. Going from 18% to 26% is eight points, which works out to roughly 44% better than the old score in relative terms. AutomationBench-AA gained six points. AA-Briefcase gained about 124 Elo, going by the approximate starting figure.

Think of a marathon runner who shaves minutes off a personal best by changing the training plan and the pacing, not by growing new legs. The body is the same size. The race looks different. That’s roughly what a post-training pass does to a model this large, and it’s why these numbers deserve attention without a victory lap.

I’d trust a release like this more than one promising a leap across the board. A three-point SciCode gain sitting next to an eight-point CritPt gain reads like real, uneven progress rather than a marketing average. Does that prove anything? No. It just sounds like how tuning actually behaves.

The competitive angle is narrower than headlines might suggest. The repository’s context points to capability gains for open-weight Chinese laboratory models outside the major US labs, with this model landing at 44 on the index. The fact pack doesn’t name rival models or their scores, so I won’t invent a leaderboard comparison.

And the caveat is the big one. These are repository-reported figures. Until independent evaluations land in the same place, 1524 Elo and 68% are what the builders say, not what the field has confirmed.

Where Ling-3.1-flash Fits

Ling-3.1-flash-1019 succeeds Ling-3.1-flash from inclusionAI, so this is an iteration, not a fresh architecture. The README frames the improvement as coming from further post-training optimizations, which fits the pattern of the numbers: moderate lifts, spread across reasoning, scientific coding and agentic work.

The design itself is worth a sentence. A mixture-of-experts model with 560B total parameters but only 25B active per token spreads its knowledge wide while keeping per-token compute lower. Pair that with a 1M-token context window and you have a model aimed at long, tool-heavy tasks, which is precisely what AutomationBench-AA and AA-Briefcase are meant to probe.

What the repository doesn’t give us is just as telling. There’s no reported controversy in the source material, and no independent verification of the new scores. We also lack a confirmed publication date for the update, so I’m not putting one on it.

That leaves a clean, narrow story. A known model family got better on the tests its authors chose to publish, and the authors say how they did it.

Three Checks Before Anyone Takes These Numbers at Face Value

First, watch for outside runs. If third-party evaluators reproduce the 44 on the Artificial Analysis Intelligence Index v4.3.2, the claim becomes a fact. If they land well below it, that gap is the real story.

Second, look at how the gains hold up on your own workloads. Benchmarks like AutomationBench-AA are proxies. A six-point rise there tells you something about agentic behavior, but not whether the model handles your messy internal tools.

Third, keep an eye on whether inclusionAI keeps shipping post-training updates to this model family. If a further revision lifts the same tests again, the pattern points to a tuning pipeline that’s still finding headroom. If it stalls, 1019 may be close to the ceiling for this base.