Ant Ling opens Ling-3.1-flash under MIT with strong coding scores

Ant Group's Ling team (inclusionAI) released Ling-3.1-flash on Hugging Face under the MIT license in October 2026. It is an upgraded version of Ling-3.0-flash and a language model for real-world tasks with a thinking mode that can be switched on or off. The model, which you can run on your own servers, is also on ModelScope and available through OpenRouter.

What the model offers

Ling-3.1-flash is a mixture-of-experts model with 560 billion total parameters, 25 billion of which are active per step. According to the model card it supports a 1 million token context. The architecture mixes linear attention layers with standard attention layers at a 7 to 1 ratio, a design used to cut compute on long inputs. Eight of its 512 experts are active at each step.

Coding and agent tests

According to the values on the Hugging Face page, the model scores 68.74 on SWE-bench Pro, which measures fixing software issues, 88.5 on SWE-bench Multilingual, 81.18 on Terminal-Bench 2.1, which measures terminal tasks, 68.7 on SkillsBench and 37.44 on Humanity's Last Exam. The card also reports gains on work, health and tool-connection tests such as GDPVal, HealthBench Professional and MCP-Atlas; the comparison chart is given as an image. The team says a coding harness called LingCode, built on Pi, may be open-sourced later.

How to run it

SGLang is the recommended path, with a reference setup of eight H200 GPUs on the card. It also runs with vLLM, and community quantized builds for llama.cpp, Ollama and LM Studio are listed. Running a model this size locally still needs very powerful hardware; trying it through a cloud provider is easier.

The scores are values shared by the Ling team, and the compared models are not named in the text. The card gives no information on language support, including Turkish. Testing it on your own code base and following independent researchers' comparisons would be useful.