Ant L3.0-flash Officially Open Source, FP8 Version Only 128GB

According to Diveplane Beating Monitor, AntBabel (inclusionAI) has officially released the Ling-3.0-flash weight, while providing both the BF16 base version and the FP8 quantized version. Both versions are released under the MIT license, and have been deployed on Hugging Face and ModelScope, available for self-deployment using SGLang or vLLM.

The BF16 base version has a weight of approximately 255GB, while the FP8 version is about 128GB, nearly halving the size. FP8 uses lower precision to store parameters, reducing storage and GPU memory requirements. In the official benchmark tests, FP8 has a maximum deviation of 1.57 points compared to BF16.

The Ling-3.0-flash model has a total of 124 billion parameters, with only 5.1 billion activated per generation. The model supports a context of 256,000 Tokens and is mainly aimed at Agent tasks such as programming, search, deep research, and tool invocation. Official evaluations show that it matches or exceeds the performance of the trillion-parameter predecessor model Ring-2.6-1T on most benchmarks.