Ant Group Open-Sources Ling-3.0-flash-VL, a 124B MoE Multimodal Model Under MIT
Ant Group’s inclusionAI lab released Ling-3.0-flash-VL under MIT—a 124B MoE multimodal model with about 5.5B active parameters—with BF16 weights on September 4 and FP8 on September 8; BlockBeats framed the open-source release on September 9.
Ant Group’s inclusionAI lab released Ling-3.0-flash-VL weights under an MIT license on Hugging Face in stages: BF16 on September 4, 2026, then FP8 on September 8. BlockBeats framed the open-source release on September 9, and the inclusionAI/Ling-3.0-flash-VL model card was last updated that day; September 9 is the announcement window, not the first weight drop. The model had been available via API before the open weights. FP4 and INT4 builds are still to come, per BlockBeats.
Ling-3.0-flash-VL is the first native multimodal checkpoint in the Ling-3.0-flash family. It keeps the sparse mixture-of-experts backbone—about 124 billion total parameters with roughly 5.5 billion activated per token—and adds image and video understanding for agent-style workflows that observe results, compare them to a goal, and correct course.
Architecture and context
The Hugging Face card describes a ViT visual encoder, a two-layer MLP projector that aligns visual features with text, and VideoRoPE for spatial and temporal order in video. A 42-layer hybrid backbone alternates KDA and Gated MLA layers at a 5:1 ratio. The card lists a context window of up to 1 million tokens; the recommended SGLang serving recipe uses YaRN to run a 256K context on four 141GB-class GPUs, or tensor-parallel 8 on 80GB cards.
Scores and serving
The model card reports a vendor Artificial Analysis Intelligence Index v4.1.1 score of 42, four points above the text-only Ling-3.0-flash figure of 38. That Index number is vendor-reported on the card and should be treated as such until independent listings catch up.
Day-zero serving paths on the card include the SGLang image lmsysorg/sglang:dev-Ling-3.0-flash-VL and a vLLM path through Ant’s inclusionAI/vllm-ling-v3 fork, with reasoning and tool-call parsers resolving to ling3. Thinking mode is enabled by default in the chat template.
Primary sources for this brief are the Hugging Face model card for inclusionAI/Ling-3.0-flash-VL (BF16 September 4, FP8 September 8, card update September 9) and BlockBeats’ September 9 open-source flash.
- #Opensource
- #AI Agents
- #Products
Author
Raj M
Contributor
AI Systems Architect is a seasoned technology leader with over 15 years of experience in the IT industry working with Fortune 500 companies. With a solid foundation in multi-agent systems, open-source LLM infrastructure, and enterprise deployment, he excels at building scalable production-grade AI platforms.