Wednesday, Oct 28 | --:--
Times of AI

Regions

China Europe India Japan Taiwan United Kingdom United States

Platforms

Alibaba Qwen Amazon Anthropic DeepMind Google Groq Meta Microsoft MiniMax Mistral Moonshot AI NVIDIA OpenAI PrismML Replit Thinking Machines xAI

Sections

Business Engineering Enterprise Frontier Labs Hardware Industry Infrastructure Models Policy Products Research Safety Security Work All articles →
Back to home
Models Open Models China

Ant Group's inclusionAI Releases Ling-3.1-flash-1019, a 560B-Parameter Open-Weight Model Under the MIT License

Ant Group's inclusionAI lab published the weights of Ling-3.1-flash-1019 on October 27 under the MIT license, along with the Ling-3.1-flash-base model. The model card describes a mixture-of-experts model with 560 billion total parameters, 25 billion active per token and a 1-million-token context window, post-trained for reasoning and agent tasks. Artificial Analysis now scores it 29 on its Intelligence Index v5.0, a different scale from the 44 on v4.3.2 that the model card cites.

Times of AI Desk 3 min read Hangzhou, China View as Markdown
A grand library with its doors flung open onto lit shelf corridors and a long blank scroll unrolling down the steps.

Ling-3.1-flash-1019 matters less for its benchmark gains than for its terms: a model of this size, with its base model alongside it, under the permissive MIT license. The catch is the hardware. Only 25 billion parameters are active per token, but all 560 billion have to be loaded, which makes this a server model.

What inclusionAI released#

inclusionAI, the AI lab of Ant Group, published Ling-3.1-flash-1019 on Hugging Face on October 27, according to Neural Brew, which also reports that the weights and the Ling-3.1-flash-base model are released under the MIT license. The Hugging Face listings for both models carry the MIT license tag.

The model card calls the release "an upgraded version of Ling-3.1-flash, with 560B total parameters, 25B active parameters per token, and a 1M-token context window." It describes the architecture as a hybrid-linear mixture of experts with 512 routed experts, eight of them active per token.

The model card's reference setup runs the model on eight Nvidia H200 GPUs, or four GB300s, using the SGLang serving framework; it also gives instructions for vLLM.

The benchmark numbers#

The model card says post-training lifted the model's AA-Briefcase score from about 1,400 to 1,524 Elo and its AutomationBench-AA score from 62% to 68%. It says the model "scores 44 on the Artificial Analysis Intelligence Index v4.3.2, up from 41 for its predecessor."

Artificial Analysis has since moved to version 5.0 of its index, which uses a different set of nine evaluations. As of 08:36 IST on October 28 (03:06 UTC), its page for the model showed a score of 29 on the v5.0 index, ranked fourth of 16 among open-weight models of similar size. The two scores are on different scales and cannot be compared directly.

The same page lists a price of $0.30 per million input tokens and $0.90 per million output tokens, and says the model generated 240 million output tokens while running the index, against a median of 99 million for comparable models.

Limits#

  • The AA-Briefcase and AutomationBench gains are inclusionAI's figures; Neural Brew notes that only the overall index score was visible on Artificial Analysis when it checked.
  • Artificial Analysis dates the model's release to October 22, while the Hugging Face repository was created on October 27; neither source explains the gap.
  • The Artificial Analysis figures above are a snapshot from October 28 and may change as the site adds models.

Sources#

Prior Coverage

Earlier Times of AI reporting on this thread.

Related Coverage

Scroll to continue reading