The model weights for Ant Group's Ling-3.1-flash have been released in the official repository under the MIT license. For developers, this means they can obtain the model files and evaluate and deploy them on their own infrastructure; the news here is the weight release, as the model had already been unveiled previously.
According to the official model card, Ling-3.1-flash uses a hybrid linear attention and mixture-of-experts architecture, with about 560 billion total parameters, about 25 billion activated parameters per token, and support for a million-token context. The developer summarizes its target tasks as general-purpose agents, search, office work, and software engineering, and describes evaluation performance in fields such as finance and materials science.
With 25 Billion Activated Parameters, Why Are Server GPUs Still Needed?
Mixture-of-experts models only call a subset of parameters into computation each time, so they can control the amount of computation per pass at a larger total parameter scale. However, the weights not invoked by the current token are still part of the complete model.
This is also key to understanding the deployment threshold: activated parameters describe the computation process and cannot be directly treated as model file size or VRAM requirements. Runtime considerations also include weight precision, context caching, concurrent requests, and inference framework overhead.
The official deployment example lists 8 H200 or 4 GB300 GPUs. The "four-card" setup here has specific server hardware requirements; it cannot be understood as any four GPUs providing the same experience.
Long Context Needs to Be Tested in Real Tasks
Take enterprise document analysis as an example: being able to input a large number of documents only solves the "fits in" problem. The system also needs to find relevant content, distinguish conflicting information, and let users trace the basis for answers.
Therefore, when evaluating a long-context model, one can repeatedly test the same question under different document lengths and different information positions to see whether the answer remains stable. Time to first token, total task time, and performance with multiple simultaneous users are also closer to the actual experience than a single capacity number.
On specifications, there is currently a discrepancy between the official model card and the accompanying SGLang documentation: the model card lists a million-token context and 25 billion activated parameters, while the latter page lists a lower context limit and different activated parameters. Therefore, this article uses the model card to describe product specifications; the specific runnable configuration still needs to be checked against the downloaded weights and framework version.
From Obtaining Weights to Building a Service
SGLang has provided a corresponding deployment solution and describes capabilities such as speculative decoding, reasoning toggles, structured tool calling, and multi-turn prefix caching. These features affect how the model is integrated into applications, but an available configuration does not mean the business process has been fully validated.
AIYXL observation: Open weights give enterprises an opportunity to control model versions and deployment environments. Whether adoption is worthwhile should also factor hardware, maintenance, task checks, and rework from errors into the cost. Running a representative set of tasks with standard answers and real tool steps as a trial is usually more informative for procurement decisions than relying directly on leaderboards.