ChainThink reports that on August 5, according to official announcements, Ant Bailing (inclusionAI) has officially released the Ling-3.0-flash weights, offering both the native BF16 version and the FP8 quantized version.
Both versions are licensed under the MIT License and are available on Hugging Face and ModelScope, supporting self-deployment via SGLang or vLLM.
Among them, the BF16 version is approximately 255 GB, while the FP8 version is about 128 GB, nearly halving the size. FP8 stores parameters with lower precision, reducing the requirements for storage and VRAM;
Among the four officially listed tests, the maximum difference between FP8 and BF16 is 1.57 points.
Ling-3.0-flash has a total of 124 billion parameters, activating 51 billion parameters per generation, supports a context length of 256,000 tokens, and is primarily designed for agent tasks such as programming, search, deep research, and tool invocation.
Official evaluations show that this model achieves or exceeds the performance of the previous trillion-parameter model, Ring-2.6-1T, on most benchmarks.
