Ling-3.1-flash
Ant Group's open 560B text model for agents and coding: 25B active, 1M context, mostly linear attention, MIT license.
In short. Ling-3.1-flash is a very large text-only AI model from Ant Group's open AI team. It is built for coding, search and office-style agent tasks, and it can read up to a million tokens at once. The MIT license lets you use it commercially with almost no conditions, but you need a multi-GPU server to run it.
Summary
inclusionAI, the open AI initiative founded by Ant Group, published Ling-3.1-flash on 13 October 2026 under the MIT license. It is a text-only reasoning model with about 560 billion parameters in total and 25 billion active per token, and it supports a 1-million-token context window.
The design is a hybrid-linear mixture of experts. Of its 72 attention layers, 63 are KDA linear-attention layers and 9 are gated multi-head latent attention layers, a 7:1 ratio. The linear layers are what keep very long context affordable. There are 512 routed experts plus one shared expert, with 8 routed experts chosen for each token. The weights are published in FP8 and take about 574 GB; a BF16 base model of about 1.1 TB is also available for research and further training.
On 27 October a second checkpoint, Ling-3.1-flash-1019, appeared with further post-training. inclusionAI reports a score of 44 on the Artificial Analysis Intelligence Index v4.3.2, up from 41 for the first version. These are the developer's own figures. The model card's reference setup is a node of eight H200 GPUs served with SGLang, and it points to a vLLM branch as an alternative. Because of the new attention layers, expect to use those specific builds rather than a stock install.
Good for
- Coding, terminal and office-automation agents that run on your own multi-GPU server
- Reading long documents and repositories, helped by the 1M-token window
- Companies that want an MIT-licensed model with no revenue or hosting conditions
- Researchers who want the base model of a hybrid linear-attention mixture of experts
Watch out for
- Text only, so it cannot read images, audio or video
- About 574 GB of FP8 weights; the card's reference setup is eight H200 GPUs, so this is not a workstation model
- Benchmark scores are inclusionAI's own and have not been independently reproduced
- The linear-attention layers need specific SGLang or vLLM builds, so a stock inference install may not work
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Estimates for the Ling-3.1-flash-1019 variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.
| Variants | Parameters | BF16 | FP8 / INT8 | 4-bit (Q4) | Fits on |
|---|---|---|---|---|---|
| Ling-3.1-flash-1019 FP8 | 560B 25B active | 1.29 TB | 644 GB | 386 GB | Big-memory desktop |
| Ling-3.1-flash FP8 | 560B 25B active | 1.29 TB | 644 GB | 386 GB | Big-memory desktop |
| Ling-3.1-flash-base BF16 | 560B 25B active | 1.29 TB | 644 GB | 386 GB | Big-memory desktop |
Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.
The license in plain words
The shortest and simplest permissive license. You can do almost anything with the model as long as you keep the copyright and license notice.
See also
Frequently asked questions
- What is Ling-3.1-flash?
- Ling-3.1-flash is a very large text-only AI model from Ant Group's open AI team. It is built for coding, search and office-style agent tasks, and it can read up to a million tokens at once. The MIT license lets you use it commercially with almost no conditions, but you need a multi-GPU server to run it.
- Can I use Ling-3.1-flash commercially?
- Yes. Ling-3.1-flash is released under MIT, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is Ling-3.1-flash open source?
- Ling-3.1-flash is released under MIT, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
- Can I run Ling-3.1-flash on my own computer?
- Only on a machine with a lot of memory. The smallest variant, Ling-3.1-flash-1019, needs about 386 GB at 4-bit precision. Desktops with 128 to 512 GB of unified memory can hold it, at modest speed.
- When was Ling-3.1-flash released?
- Ling-3.1-flash was released by inclusionAI (Ant Group) on October 13, 2026.
- How long a context window does Ling-3.1-flash have?
- Ling-3.1-flash supports a context window of up to 1M tokens.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- Ling-3.1-flash-1019 model card (Hugging Face) huggingface.co
- Ling-3.1-flash model card (Hugging Face) huggingface.co
- SGLang cookbook: Ling-3.1-flash docs.sglang.io
Release date and facts last verified November 8, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| DeepSeek V4.1 Flash DeepSeek | 552B 16B active | 1M | MIT (Permissive) | 561 GB GPU server | |
| GLM-5.3-Flash Z.ai (Zhipu AI) | 320B 18B active | 1M | MIT (Permissive) | 221 GB Big-memory desktop | |
| MiniMax M3.1 Flash MiniMax | 428B 23B active | 1M | MiniMax Community (Conditional) | 277 GB Big-memory desktop | |
| Ming-Image 0.1 Design inclusionAI (Ant Group) | 6.2B | – | MIT (Permissive) | 80 GB Workstation |