Ling 3.1 Flash: InclusionAI releases 560B open-weight model under MIT license

What's going on

InclusionAI, the AI lab of Ant Group (best known for Alipay), has released Ling 3.1 Flash, a large language model, as an open-weight model. "Open weight" means you can download the trained weights – roughly, the settings the model learned during training – and run the model yourself. According to the Model Card, which Techapple's report draws on, the weights were released on October 16, 2026 under the permissive MIT license, in fp8 format (8-bit floating-point numbers that use less memory than 16-bit). Ready-made launch recipes are also available for the inference frameworks SGLang and vLLM, the programs used to run models like this on servers.

Some context: the model wasn't new. Ant announced it on September 30, 2026 and initially offered it only in its own product and through its own API. An OrcaRouter report written in mid-October explicitly noted that there were no weights, no Model Card and no license. So the current release follows the plan Ant itself had laid out.

Architecture: big, but frugal

Ling 3.1 Flash is a Mixture-of-Experts (MoE) model. Think of it as a team of specialists: the model contains a great many areas of expertise ("experts"), but only a few are consulted for each word (token). That saves compute.

• Parameters: about 560B in total, roughly 25B active per token – a ratio of about 1 to 22.
• Experts: 512 routed experts plus 1 shared expert that always contributes. According to OrcaRouter (citing Ant's announcement), each token picks 8 of them.
• Attention: a hybrid design of 63 KDA layers and 9 Gated-MLA layers. KDA stands for Kimi Delta Attention, a "linear" attention whose compute cost grows only proportionally with text length rather than quadratically. The Gated-MLA layers keep a global view of the entire text. That works out to 7 linear layers for every 1 global layer.
• Context: the model is designed for 1 million tokens. That's roughly the volume of several thick books, all "in view" at once.

For comparison, the predecessor Ling-3.0-flash has 124B parameters, with 5.1B active. Ling 3.1 Flash is therefore about 4.5 times as large and activates roughly five times as many parameters per token.

Performance, according to the vendor

The official figures cover two coding benchmarks:

Both tests check whether a model can solve real software engineering tasks. The sources contained no independent re-measurements. Ant's announcement also cites long-horizon results, such as a Lua-to-x86-64-ELF compiler built in about 17 hours without interruption that passed 178 of 182 tests (97.8%). In a Pyright optimization run of about 10 hours, the model reached 1.432x speed with all 2,279 tests passing. These are also vendor numbers on self-defined tasks, and the competitors' figures are expressly just "listed" there, not a head-to-head comparison under identical conditions.

The model is positioned for long, real-world tasks: office work from a vague request to a finished result, software development that runs for hours, and agents that call tools one after another and correct themselves. It handles text only; for images, Ant points to Ling-3.0-flash-VL.

Context: 1 million or 256K?

During the test phase, the service ran with 256K tokens. Ant's plan was to open the full 1M window once the free phase ended and to release the model as open source at the same time. Whether the newly released weights make the full 1M usable isn't clear from the sources. The memory costs of such a long context also remain an open question. If in doubt, plan for 256K.

What you need to try it

This isn't a model for your home PC. The fp8 weights alone are likely several hundred gigabytes (560B parameters ≈ 560 GB at 8 bits). Even with 4-bit quantization, you're looking at roughly 280 GB plus room for the context. A realistic setup is a multi-GPU server with several 80 GB cards or a cloud cluster, plus correspondingly large RAM and storage. It won't run in any sensible way on a Mac, a regular CPU or the free tier of Google Colab. Whether GGUF versions or an Ollama entry exist isn't stated in the sources. For smaller hardware, the predecessor Ling-3.0-flash or Ling-3.0-tiny makes more sense.

Our take

An MIT-licensed model of this size with strong coding scores is interesting for companies and research groups with their own GPU infrastructure, because they can run and adapt it without license hurdles. The release lands amid a wave of cheap, open, agent-capable models, which according to Techapple also includes DeepSeek V4.1-Flash-1017 and Tencent's Hy Memory 2.0. A few caveats: all benchmark figures come from the vendor, the charts in Ant's announcement contain no citable table of individual values, and the sourcing isn't entirely clear – an older report still said the release wasn't happening. We rely on the Model Card as reported by Techapple, so take a look at the repository yourself before deploying. Because of its size, the model has little practical use for home users.