Two scores, both printed by the same evaluator

The number on the model card is 44. It is not invented, and the card is unusually clear about where it comes from: the Artificial Analysis Intelligence Index, version 4.3.2, up from 41 for the model this one replaces. The gains around it point the same way. According to inclusionAI's model card, AA-Briefcase rose from approximately 1400 to 1524 Elo and AutomationBench-AA from 62% to 68% relative to the predecessor. These are developer-reported comparisons, not independently reproduced results.

Open the evaluator's own page for the same model today and the headline reads 29. That page is explicit about its version too: Intelligence Index v5.0, built from nine evaluations, several of them revised or new. The 29 places the model above the median of 23 for open-weight models of similar size, which is a respectable result stated plainly. What it is not is a decline from 44. The two numbers come from different tests with different components, and neither can be ranked against the other.

This is the gap the release will travel through. A score quoted without its version is a score quoted without its test, and 44 is the version that fits in a headline. Readers comparing this model against anything else should compare only within one version: on v5, the current one, Ling-3.1-flash-1019 sits at 29.

What actually shipped

Two checkpoints, and they do different jobs. The -1019 model is the finished, post-trained release: the one with the benchmark claims, the reasoning behaviour and the API pricing Artificial Analysis lists at $0.30 per million input tokens and $0.90 per million output. The base checkpoint belongs to its predecessor, Ling-3.1-flash, and it is raw material by design. Its card offers it for continued pretraining, domain adaptation and post-training research, and says in plain terms that it is not recommended as-is for direct chat, for safety-critical work without further alignment and evaluation, or for production without validation.

Releasing a 560B-parameter base checkpoint alongside the post-trained model gives researchers access to a different starting point: weights intended for continued training, adaptation and post-training experiments rather than direct end-user deployment. A 560B-parameter foundation checkpoint, evaluated by its maker on 47 benchmarks against the base versions of DeepSeek-V4-Pro and MiMo-V2.5-Pro, gives other labs a starting point they can actually train from. Whether they do is a separate question, and it will be answered by downloads and derivatives, not by the card.

The serving bill, and two checks before you plan around it

Open weights do not mean small hardware. The card's recommended SGLang setup runs on eight H200 GPUs, with four GB300s as the alternative, and the vLLM route asks for a branch build from source with remote code trusted and model-specific parsers switched on. This is a model for an organisation's cluster or a rented one. Artificial Analysis reports that the model generated 240 million output tokens during its index evaluation, compared with a 99 million median for similarly sized open-weight models. That suggests higher aggregate output-token usage under this evaluation, but it does not establish the cost or token count of an individual workload.

Two details deserve a check before anyone builds a plan on the card. Both inclusionAI's -1019 model card and Artificial Analysis list a one-million-token context window. The -1019 repository's own configuration, read directly for this article, sets max_position_embeddings at 262,144, and its rope_scaling entry is null. Teams planning long-context deployment should verify the supported serving settings before promising a million-token workload. The -1019 model card's SGLang quickstart also uses inclusionAI/Ling-3.1-flash as its model path rather than the -1019 repository. Users intending to serve the upgraded checkpoint should verify the repository identifier before copying that example. Neither detail undermines the release. Both say the documentation trail is younger than the weights.

Sources

  1. [1] inclusionAI, "Ling-3.1-flash-1019," Hugging Face model card, read October 28, 2026 (primary source: specifications, architecture, benchmark gains as reported by the developer, 44 on Intelligence Index v4.3.2, serving quickstart; the repository's licence metadata, config.json and commit history were also read directly for this article)Read source
  2. [2] inclusionAI, "Ling-3.1-flash-base," Hugging Face model card, read October 28, 2026 (primary source: base checkpoint scope, BF16/WSM release details, intended and not-recommended uses, MIT licence statement, 47-benchmark vendor evaluation)Read source
  3. [3] Artificial Analysis, "Ling 3.1 Flash 1019 — Intelligence, Performance & Price Analysis," read October 28, 2026 (evaluator's own page: 29 on Intelligence Index v5.0, v5.0's nine component evaluations, release date of October 22, 2026, pricing, output-token count, MIT licence listing)Read source
  4. [4] Victor Grösser, "Ant Group just open-sourced a 560B agent model under MIT, and the base checkpoint too," Medium, October 28, 2026 (reporting of the October 27 Hugging Face publication of the -1019 weights and the base checkpoint two days earlier)Read source
  5. [5] AlexTech, "Ling-3.1-flash gains on benchmarks, open weights to follow," October 22, 2026 (contemporaneous record that on October 22 the open weights were listed as coming soon, with no release date given)Read source