Ant Group Open Sources Ling-3.1-flash With a Million Token Context
Ant Group's inclusionAI has published the weights of its Ling flash model on Hugging Face under the MIT licence, completing a plan the team set out when the model was announced. The release carries a 560 billion parameter mixture of experts design with 25 billion parameters active for each token, and a context window that reaches one million tokens.
The launch came in two stages. The model was announced with a free API trial that ran for two weeks and was capped at a 256K token context, and the team said at the time that it would open the full window and publish the weights once the free period ended and the service became paid. The trial closed and the weights went up alongside it.
But the architecture is the reason a model of this size can be served at all. It uses a hybrid linear mixture of experts design with 63 linear attention layers and nine Gated MLA layers in a seven to one ratio, four dense layers, 512 routed experts plus one shared expert, and eight activated experts. Only a fraction of the network engages for each token, which is how the active count stays at 25 billion.
Ant says the model improves on general purpose agents, search, office work and software engineering. The model card reports strong results across coding and agent evaluations, naming FrontierSWE and SkillsBench among them alongside HealthBench Professional, and it cites gains in finance and materials science as well.
Deployment support arrived with the weights rather than after them. The card recommends SGLang, with a dedicated image for speculative decoding, and gives a reference server setup on eight H200 GPUs or four GB300 GPUs. A vLLM recipe is provided, and the model is listed on ModelScope and OpenRouter.
But the most revealing part of the release may be the tooling built around long jobs. The card introduces LingCode, a harness built on the open source coding agent Pi and tuned for work that runs for hours, involves hundreds of tool calls and needs repeated context compaction. Ant says it is considering open sourcing LingCode in the future.
The model is also pitched at third party coding frameworks, with the card naming Claude Code, Kilo Code, Qwen Code, Hermes Agent and OpenClaw as environments it works well inside. That is a distribution strategy as much as a technical claim, since developers already working in those tools can try the model without rebuilding their setup.
The release continues a rapid cadence from Ant's AI unit. It follows the earlier flash model in the same family and a multimodal version of it, both shipped within the past few months, and the licensing terms point to a group competing for developers who prefer open weights to a hosted API.
So the significance lies less in any single benchmark than in the terms of the release. A model of this scale, with a million token window, a permissive licence and deployment recipes attached, gives teams a self hosted option that was not available in this class a year ago. Whether it performs as described outside Ant's own testing is the question the model card cannot answer.
Ant Group's inclusionAI published the Ling flash weights on Hugging Face under the MIT licence
The model has 560 billion total parameters, 25 billion active per token and a one million token context window
It uses a hybrid linear mixture of experts architecture with 512 routed experts and eight activated experts
Recommended deployment is SGLang on eight H200 GPUs or four GB300 GPUs, with a vLLM recipe also provided
Ant also introduced LingCode, a harness for long coding jobs, and says it may open source it later
Source: Pandaily
