Ling-3.1-flash-1019

inclusionAI's MIT open-weight Ling upgrade for reasoning and agent work.

Also known as Ling-3.1-flash, Ling-3.1-flash-base

What it does

Ling-3.1-flash-1019 is an open-weight upgrade of Ling-3.1-flash from inclusionAI. The Hugging Face model card lists 560 billion total parameters, 25 billion active per token, and a 1 million token context window. The published config.json sets max_position_embeddings to 262,144. The card describes a hybrid-linear mixture of experts, with 63 KDA layers, 9 gated MLA layers, 512 routed experts, 8 active, and 1 shared expert, and says post-training improved reasoning and agentic work. It reports an Artificial Analysis Intelligence Index v4.3.2 score of 44, up from 41. Weights are MIT on Hugging Face and ModelScope. The Hub API lists 572,120,680,768 parameters, mostly FP8, and a repository created on October 27, 2026. The card's SGLang example still passes --model-path inclusionAI/Ling-3.1-flash. A separate checkpoint, Ling-3.1-flash-base, is the BF16 base of Ling-3.1-flash for further training, also under MIT, with a repository created on October 25, 2026. Neither card lists a hosted price.

Who is Ling-3.1-flash-1019 for?

Teams running an open-weight model for reasoning and agent workflows, or starting further training from the base checkpoint.

Tags

• open-weight
• agents
• coding