ChainThink News, August 5: According to official announcement, InclusionAI's Ling-3.0-flash model weights are now officially released, available in both BF16 native and FP8 quantized versions.

Both versions are licensed under MIT and have been deployed on Hugging Face and ModelScope, supporting self-deployment via SGLang or vLLM.

The BF16 weights are approximately 255GB, while the FP8 version is around 128GB—nearly halved in size. The FP8 variant stores parameters at lower precision, reducing storage and GPU memory requirements;

In four official benchmarks, the maximum performance gap between FP8 and BF16 was 1.57 points.

Ling-3.0-flash has a total of 124 billion parameters, activating 5.1 billion parameters per generation, with support for up to 256,000 tokens of context, primarily targeting Agent tasks such as coding, search, deep research, and tool calling.

Official evaluations indicate that the model achieves or exceeds the performance of its predecessor, the trillion-parameter Ring-2.6-1T model, across most benchmarks.