Ling 3.0

InclusionAI's third-generation open-weight reasoning family replaces full attention with a native KDA–MLA hybrid. The flagship Ling-3.0-flash is a 124B-parameter sparse MoE activating 5.1B parameters per token: 35 Kimi Delta Attention layers alternate with seven gated Multi-head Latent Attention layers, while 512 routed experts activate eight per token. It was pretrained with this architecture rather than converted after the fact, supports a 256K context, and is released under the MIT license.

Post-training used more than 10,000 interactive coding, research, and general-agent environments. The model card reports integration with SGLang HiCache and Mooncake's shared L3 cache, reducing time to first token by 60–80% for long-input agent workloads. Artificial Analysis scores the reasoning checkpoint at 38 on Intelligence Index v4.1.1.

Ling-3.0-tiny followed on August 11 with 7.9B total / 1.3B active parameters, a 3:1 KDA–MLA stack, 128 routed experts, and the same 256K context. Its FP8 checkpoint runs at roughly 86–90 tokens/s on an M4 Pro MacBook according to InclusionAI; AA scores it at 25.