The Pulse
InclusionAI Posts Ling-3.1-Flash-1019 With 560B Parameters
InclusionAI has posted Ling-3.1-Flash-1019, a 560-billion-parameter model with 25 billion active parameters per token and a one-million-token context window. Its Hugging Face release lists an MIT license and reports gains on several reasoni
AI.info Team · Written with AI · fact-checked ·
InclusionAI’s Ling-3.1-Flash-1019 has 560 billion total parameters, but activates 25 billion for each token it processes. The company posted the model weights on Hugging Face under the MIT license, alongside a model card describing a one-million-token context window and improvements over the earlier Ling-3.1-Flash.
The October 27 release is a revised checkpoint, not a new model family. InclusionAI says post-training changes improved reasoning and agent performance; the scores in the model card are the company’s own reported results, not independent confirmation of those claims.
560 billion parameters, 25 billion active
The model uses a hybrid-linear mixture-of-experts architecture. Its specification lists 512 routed experts and one shared expert, with eight experts active for each token. That design separates the model’s total parameter count from the smaller subset used during an individual computation.
The Hugging Face repository lists the model at 574 GB and divides its weights among 54 Safetensors files. The card identifies BF16, F8_E4M3 and F32 tensor types. The total parameter count and the active-per-token count describe different things: 560 billion is the full model, while 25 billion is the amount activated per token.
InclusionAI reports higher benchmark scores
InclusionAI says the updated model scores 1,524 Elo on AA-Briefcase, compared with approximately 1,400 for its predecessor. It reports AutomationBench-AA rising from 62% to 68%, CritPt from 18% to 26%, and SciCode from 54% to 57%. The card also gives Ling-3.1-Flash-1019 a score of 44 on the Artificial Analysis Intelligence Index version 4.3.2, up from 41 for the previous version.
“These gains highlight the model's improved capabilities in complex reasoning, multi-step task execution, and agentic workflows across diverse real-world scenarios.”
InclusionAI, model card
Those figures come from InclusionAI’s description of its release. They provide comparisons with the prior Ling version, but the model card’s benchmark claims do not by themselves establish how the model will perform on a particular team’s tasks.
Public weights, substantial serving requirements
The MIT license makes the repository’s terms clear, and the files are publicly available to download. Running the model is a separate matter: InclusionAI’s SGLang instructions give an eight-H200 setup as a server example, while also listing four GB300s as an alternative. The model card includes a vLLM serving recipe as well.
The release page says Ling-3.1-Flash-1019 is not deployed through Hugging Face Inference Providers. Users seeking to test it therefore need an appropriate local or hosted environment rather than a ready-to-use provider endpoint on that page.
A longer context, with practical trade-offs
InclusionAI lists a one-million-token context window, a feature intended to let the model process much longer inputs than systems with shorter context limits. That specification does not, on its own, establish how well the model retains or uses information across a full-length prompt. The 574 GB repository size and the example multi-GPU configurations make the immediate trade-off plain: the weights are accessible, but operating the full model calls for considerable compute and storage.
Ling-3.1-Flash-1019 gives developers a licensed checkpoint to inspect and run, and InclusionAI has published benchmark comparisons with its prior model. The open questions for adopters are whether those reported gains hold on their own evaluations and whether the hardware burden fits their deployment plans.