560B total / 25B active (hybrid-linear MoE)
Up to 1M tokens advertised (262,144-token checkpoint configuration)
MIT
MIT-licensed open-weight upgrade to Ling-3.0-flash and native hybrid-reasoning text-generation model aimed at general agents, search, office productivity, software engineering, healthcare, finance, and materials science. Its hybrid-linear Mixture-of-Experts backbone has 72 transformer layers arranged as 63 Kimi Delta Attention layers and nine Gated Multi-head Latent Attention layers in a repeating 7:1 pattern, with four initial dense layers, 512 routed experts selecting eight per token, one shared expert, 64 attention heads, a 4,096-dimensional hidden state, and a 157,184-token vocabulary. inclusionAI reports 560B total parameters and 25B activated per token; Hugging Face's safetensors metadata inventories approximately 572.1B serialized parameters, while the linked SGLang cookbook estimates roughly 17B active from its expert accounting. The released checkpoint uses blockwise FP8 E4M3 weights with dynamic activation quantization and includes one NEXTN multi-token-prediction layer for speculative decoding. Its 54 downloadable safetensors shards total approximately 573.8 GB. The model card advertises a context window up to 1M tokens, although the checkpoint configuration and verified SGLang deployment documentation specify 262,144 positions, so the advertised million-token operating regime is not encoded in the published configuration. Thinking is enabled by default but can be disabled, and the supplied chat template preserves reasoning across turns while supporting structured tool calls. inclusionAI reports 1,673 Elo on GDPVal-AA V2.1, 75.16 on FrontierSWE, and 65.35 on HealthBench Professional. The release also introduces LingCode, a Pi-derived harness for multi-hour coding tasks, hundreds of tool calls, and repeated context compaction, but inclusionAI has not yet open-sourced that harness. Official deployment guidance covers SGLang and vLLM, including tensor and expert parallelism, prefix caching, structured tool and reasoning parsers, and NEXTN speculative decoding; the linked SGLang recipe reports 251.2 output tokens per second at batch size one on eight H200 GPUs.