On September 9, Ant Group's inclusionAI open-sourced Ling-3.0-flash-VL, the first natively multimodal model in its Ling line, on Hugging Face and ModelScope: 124B total parameters with only 5.5B activated per token, image and video input, a 256K-token context window, and BF16 plus FP8 weights available immediately, with FP4 and INT4 builds planned to follow.

Architecture: sparse MoE under a hybrid-attention backbone

The model card describes a straightforward stack: a ViT visual encoder extracts image and video features, a two-layer MLP projector aligns them with text representations, and VideoRoPE encodes both spatial positions and temporal order — supporting event localization, long-video question answering, and video clip editing. The language trunk is a 42-layer hybrid backbone alternating KDA and Gated MLA layers at a 5:1 ratio for efficient long-context processing, wrapped in a sparse MoE that holds 124B of total capacity while activating 5.5B per token. Vision here is not a bolted-on captioning plugin; the card frames it as integrated into the full loop of understanding, reasoning, acting, and verification.

Benchmarks: the official citation and the independent eval disagree

The official card claims 42 on the Artificial Analysis Intelligence Index v4.1.1, four points above the text-only Ling-3.0-flash at 38 — Ant's argument being that joint multimodal training improved text ability as well. But Artificial Analysis's own evaluation post puts the model at 25, while placing it on the Intelligence-vs-Active-Parameters Pareto frontier: among models of similar total size, Qwen3.5 122B A10B scores 16 and Mistral Medium 3.5 (high) scores 15. The two numbers most likely reflect different index versions — which is precisely the reminder that vendor-cited benchmarks deserve a version check before you quote them.

AA's independent evaluation also surfaces the weaknesses no launch deck highlights: a 22% hallucination rate (restrained next to Inkling Small's 63%), but only 14% accuracy on AA-Omniscience factual recall; 16% on AutomationBench-AA business-workflow automation; and 0% on Terminal-Bench v4.0, the harder terminal-use benchmark. The see-and-act story is real; the reliable-digital-worker part is not yet.

Deployment reality: 5.5B active is not 5.5B of VRAM

The number people misread: active parameters reduce per-token compute, not memory. You still host all 124B of weights. The card's reference configuration for 256K context (YaRN extension) is 4x 141GB-class GPUs (H20-3e / H200, or B300 / GB300 nodes); 80GB H100/H800 cards scale out to 8-way tensor parallelism. Two serving paths: the official SGLang Docker image, or inclusionAI's vllm-ling-v3 fork for vLLM.

So what

Open-multimodal competition has moved from "can describe an image" to "can close the loop on real work" — GUI operation, front-end restoration, and medical-report reading are the three scenario families the release names. The 5.5B-active inference economics aim squarely at the flash tier, the highest-volume pricing segment of closed-model vendors. But for practitioners, the signal worth remembering is not the official 42 — it is the independent 0% on Terminal-Bench v4.0. Multimodal agents still have deep water ahead.

References: model card · Artificial Analysis eval · Pandaily