124B total / 5.5B active (MoE)
Up to 1M tokens (131,072 native configuration; official 262,144-token YaRN recipe)
MIT
Official open-weight native multimodal extension of Ling-3.0-flash, adding image and video inputs to the model's text generation, reasoning, tool use, and agentic capabilities. inclusionAI positions vision as part of the complete understanding, planning, acting, and verification loop rather than only a perception interface, with reported support for complex layouts, charts, documents, long-video question answering, temporal event localization, visual verification, software interfaces, and visually grounded action generation. The visual stack uses a 27-layer ViT encoder with a 1,152-dimensional hidden state, 16 attention heads, 16-by-16 patches, temporal patching, and a two-layer MLP projector, while VideoRoPE represents both spatial position and temporal order. The 42-layer language backbone alternates 35 Kimi Delta Attention layers with seven Gated Multi-head Latent Attention layers in a repeating 5:1 pattern. Its sparse Mixture-of-Experts architecture contains 512 routed experts, selects eight per token, begins with two dense layers, and activates a marketed 5.5B of 124B total parameters per token. The model card advertises context up to 1M tokens, although the checkpoint configuration specifies 131,072 native positions and the published SGLang recipe demonstrates 262,144 tokens using YaRN factor 2. Thinking is enabled by default but can be disabled per request, and the supplied chat template supports structured tool calling and image or video message content. inclusionAI reports a score of 42 on Artificial Analysis Intelligence Index v4.1.1, four points above the text-only Ling-3.0-flash result reported on the same version. The release includes a roughly 250 GB BF16 checkpoint split across 64 safetensors shards and a same-day official FP8 checkpoint, with deployment guidance for SGLang and inclusionAI's Ling-compatible vLLM fork.