Anyone still downloading the latest Ling-3.1-flash may want to wait: InclusionAI is preparing a new checkpoint. Ant Group's lab has announced Ling-3.1-flash-1019, which according to Artificial Analysis scores 44 points on the Intelligence Index, up from 41 for the previous version.
The sharpest gain concerns agentic capabilities, meaning the model's ability to complete multi-step tasks using external tools. On AutomationBench-AA the score rises from 62% to 68%, while on AA-Briefcase the Elo goes from roughly 1400 to 1524. These figures are reported by the company, which presents them as improvements over the current base.
Image: inclusionAI — https://huggingface.co/inclusionAI/Ling-3.1-flash
The 1019 version is not yet available to everyone. Ant Ling lists both API access and the release of open weights, meaning the model parameters anyone can download and run on their own infrastructure, as "coming soon". No release date has been given; for updates, the company points to its official Discord server.
The starting model is already known to AlexTech readers. Ling-3.1-flash was released at the end of September with 560 billion total parameters, about 25 billion active per token, and a one-million-token context window; in October its weights were made available for development on proprietary infrastructure. Its predecessor, Ling 3.0 Flash, had already halved the hallucination rate to 44% in the AA Omniscience tests, down from 97%.
Image: inclusionAI — https://huggingface.co/inclusionAI/Ling-3.1-flash
The strategy mirrors what other Chinese labs are doing: first an API version, then an open-weight distribution, with intermediate updates measuring progress on independent benchmarks. For those running open-weight models locally, the practical question is when to download the new checkpoint and whether the agentic gains hold up outside the company's own tests.
No comments yet. Be the first!