124B total / 5.1B active (MoE)
262,144 tokens
MIT
Next-generation native hybrid-reasoning language model from Ant Group's inclusionAI, designed for efficient production-scale agents and released with only 5.1B of its marketed 124B parameters active per token. Ling-3.0-flash uses a 42-layer hybrid-linear Mixture-of-Experts architecture comprising 35 Kimi Delta Attention layers and 7 gated Multi-head Latent Attention layers in a repeating 5:1 pattern, with two initial dense layers followed by a fine-grained 1/64 sparse MoE. The architecture contains 512 routed experts plus one shared expert, selects eight routed experts per token, and uses a 2,560-dimensional hidden state with 32 attention heads. Its pretraining context schedule progresses from 8K to 32K and finally 256K, while the published configuration supports 262,144 tokens. The model combines thinking and non-thinking behavior in one checkpoint, with thinking enabled by default, and provides native structured tool calling for multi-turn agent workflows. inclusionAI says its agent-focused training used more than 10,000 interactive environments spanning coding, general agent tasks, and deep research, alongside a hierarchical SGLang HiCache and Mooncake caching design intended to reduce redundant computation in long-running interactions. Official reported results include 93.2 on AIME 2026, 87.0 on HMMT February 2026, 72.4 on SWE-Bench Multilingual, 56.6 on SWE-Bench Pro, and 22.7 on Humanity's Last Exam. The release includes official BF16, FP8, FP4, and INT4 repositories under the MIT license, custom BailingMoeV3 model code, MTP/NEXTN speculative decoding support, and deployment instructions for SGLang and inclusionAI's Ling-specific vLLM branch.