Improved version about a week later
inclusionAI, the AI unit of China's Ant Group, on Oct. 22 announced "Ling-3.1-flash-1019," an improved version of its large language model "Ling-3.1-flash." About a week after releasing the original version in mid-October, the company strengthened reasoning and agent tasks through additional training. The company's X account @AntLingAGI made the announcement.
According to the company, its score on the intelligence index of third-party U.S. firm Artificial Analysis was 44, up 3 points from 41 for the original version. Its scores also improved on evaluations measuring agent tasks, the company said. At the time of the announcement, it said the API and weights would be released "soon."
| Version | Intelligence index |
|---|---|
| Ling-3.1-flash (mid-October) | 41 |
| Ling-3.1-flash-1019 (Oct. 22) | 44 |
560 billion parameters, 25 billion in use
The model uses an architecture that switches between only some of its many expert components. It has 560 billion total parameters, of which only 25 billion are used per pass. It uses 8 of its 512 experts. The weights were released on Oct. 27 on Hugging Face of the U.S. under the MIT license, which also permits commercial use.
On Artificial Analysis' page, the context window is 1 million tokens, and pricing is $0.30 for input and $0.90 for output per 1 million tokens. That is cheaper than the median for open models of the same size ($0.43 for input and $1.68 for output). However, the page rates the amount of text the model produces before answering as on the high side.
Original version runs on half the hardware
The original Ling-3.1-flash compressed part of its weights at a finer level, reducing its storage size from about 534GB to 292GB. In the company's tests, the number of NVIDIA graphics processors needed to run it fell from 8 to 4, it said. inclusionAI on Oct. 26 also released the weights of "LBS-27B," a model with no description, under the Apache 2.0 license, but did not state its uses or performance. Ant Group on Oct. 19 also released the weights of its voice model "Ming-3.0," and it has been releasing open models one after another.
