Ling-3.1-flash aims to combine power, speed, and deployment flexibility.

Ant Group releases Ling-3.1-flash, an open-source 560-billion-parameter AI model designed for code, agents, and interface creation.

Transforming a rough brief into a mobile interface, producing a web page, or building a manipulable 3D prototype in a browser: Ling-3.1-flash aims to bring reasoning closer to practical creative and development uses. Developed by inclusionAI, Ant Group's research team, the model is becoming accessible as open source under the MIT license. It combines a 560-billion-parameter architecture, a context window of up to one million tokens, and particularly fast performance in Artificial Analysis evaluations. A large model that activates only a portion of its capabilities

Ling-3.1-flash relies on a hybrid Mixture-of-Experts (MoE) architecture. Out of its 560 billion parameters, roughly 25 billion are activated for each processed token. The model thus distributes computation among specialized components, rather than mobilizing all of its resources at each step.

The architecture combines different attention mechanisms to process large contexts while limiting computational requirements. Its announced context window reaches one million tokens, a volume suited for analyzing long documents, software projects, or agent interaction histories.

Ant Group also worked on an optimized version to reduce memory requirements. According to its tests, the size of the weights drops from approximately 534 GB to 292 GB, thanks to a combination of less memory-intensive numerical formats. A setup previously run on eight Hopper GPUs was able to be adapted to four GPUs.

This optimization facilitates deployment on private infrastructures, without making Ling-3.1-flash a model intended for ordinary personal computers. From web pages to 3D prototypes

The model targets several professional uses, including software development, interface creation, and task automation.

From text prompts, Ling-3.1-flash can produce React Native screens, editable SVG illustrations, websites, or interactive 3D prototypes. These creations can then be reviewed and adjusted by teams.

In Design Arena Open Weights, Ant Group announced a third-place ranking for React Native interfaces and 3D creations, fourth for SVG, and fifth for websites. However, these rankings depend on the models evaluated and the specific criteria for each category.

The appeal for designers and developers lies particularly in the production of actionable elements rather than simple visual representations. An interface component or a 3D scene can be incorporated into a project, modified, and integrated into a development workflow. Around 210 tokens per second

Speed is another key selling point of Ling-3.1-flash. On Artificial Analysis, the model achieves an Intelligence Index of 41 points, with a generation speed measured around 209 tokens per second.

Its weighted average price is estimated at around $0.19 per million tokens. This amount corresponds to the Artificial Analysis comparison methodology and not to a universal rate applicable to all providers.

These figures help position Ling-3.1-flash along two distinct dimensions: its general performance and its execution speed. High throughput can be useful for agents performing many sequential operations, but on its own, it does not guarantee the success of a complex project.

Ant Group also claims progress on several professional benchmarks, notably GDPVal-AA V2.1, FrontierSWE, Terminal-Bench 4.0, and tool-use evaluations. LingCode for extended development tasks

In parallel, inclusionAI presented LingCode, an execution environment based on the open-source coding agent Pi.

It is designed to support workflows that can last several hours, requiring hundreds of tool calls and multiple context compression operations to continue the task.

Ling-3.1-flash is also compatible with existing environments such as Claude Code, Qwen Code, Hermes Agent, or OpenClaw. Its support by SGLang and vLLM allows developers to run it on their own servers.

LingCode could be released as open source later, but Ant Group has not yet confirmed this release. An open model, but a demanding infrastructure

The Ling-3.1-flash weights are available on Hugging Face and ModelScope under the MIT license. Companies can therefore study, adapt, and deploy the model in their own environments, provided they have the necessary hardware resources.

This release avoids relying exclusively on a proprietary API to leverage a large-capacity model. In return, server management, inference costs, and optimizations remain the responsibility of those who choose to host it.

With Ling-3.1-flash, Ant Group strengthens its presence in open models designed for professional agents. Its interface and prototype generation capabilities represent an interesting angle for creative fields, but their true value will depend on the quality of the generated code, its ease of modification, and its reliability in real-world projects.