Ling 3.0 Flash VL

Ant Group's inclusionAI lab released Ling-3.0-flash-VL on 10 September 2026, a 124-billion-parameter open-weights mixture-of-experts model that activates 5.5 billion parameters per token and handles text, images, and video within a 262,144-token context window. The weights ship under the permissive MIT licence, with fp The vision pipeline combines a ViT encoder with a small two-layer projector and a VideoRoPE positional scheme that encodes temporal position, enabling queries that span long videos and retrieval of specific moments. The backbone alternates Kimi Delta Attention and Gated MLA at a 5:1 ratio across 42 layers, which the la

Artificial Analysis's dedicated model page lists Ling-3.0-flash-VL as an open-weights InclusionAI model released in September 2026 with an AA Intelligence Index v4.1.1 score of 25, placing it well above the median for its open-weights size class. It records 144.6 output tokens per second (above the 134 median), 160 mil Technical specifications confirm a 124B-total-parameter MoE architecture with 5.5B active parameters per token and reasoning enabled on this variant. The page's comparison summary characterizes the model as "amongst the leading models in intelligence" within its open-weights size class and "faster than average," while

InclusionAI has released Ling-3.0-flash-VL, a vision-language model extending its Ling 3.0 Flash series into multimodal territory. Built on a mixture-of-experts architecture, the model pairs image understanding with text generation and ships under a permissive MIT license that permits commercial use and redistribution. The flash branding signals a focus on speed and efficiency through sparse activation. The release supports vision-language plus text generation modalities using an MoE design that activates only a fraction of total parameters per token. MIT licensing lowers barriers for developers building products on top of the model, making it a practical option for vision-language applications. As the first vision-language entry in the Ling 3.0 Flash family, it broadens an already active open-weights lineup.

Ling-3.0-flash-VL is listed in a third-party model catalog as an open-source multimodal variant from the Ling-3.0 series, released under the MIT license and available with downloadable weights. The catalog records a Mixture-of-Experts architecture with 124 billion total parameters and 5.5 billion active per token, supporting text, vision, and code modalities. The same catalog entry reports a context window of 262K tokens with a maximum output of 33K tokens, positioning Ling-3.0-flash-VL as a long-context multimodal model. No benchmark scores have been recorded yet, and the catalog explicitly notes that vendor-reported figures remain unverified until independently tested.

InclusionAI introduced Ling-3.0-flash-VL as a native multimodal model extending Ling-3.0-flash with visual information integrated into understanding, reasoning, acting, and verification. It carries 124B total parameters with only 5.5B activated per token, supports image and video inputs, and offers a context window of up to 256K tokens for powerful multimodal reasoning. The architecture pairs a ViT visual encoder with a two-layer MLP projector and uses VideoRoPE for spatiotemporal encoding across video tasks. A 42-layer hybrid backbone alternates KDA and Gated MLA layers at a 5:1 ratio, while a sparse MoE design balances capability with inference efficiency. Ling-3.0-flash-VL scores 42 on the Artificial Analysis Intelligence Index v4.1.1, a 4-point gain over its text-only sibling.

Crafiq's model profile names Ling-3.0-flash-VL by inclusionAI, released September 2026, as the first vision-integrated model in the Ling-3.0-flash family, supporting text, image, and video inputs with text outputs for applications such as document intelligence, medical report interpretation, and agentic workflows with Visual processing is described as powered by a ViT encoder paired with a two-layer MLP projector that aligns visual features with the text embedding space, and VideoRoPE is credited with encoding both spatial and temporal frame positions to enable event localization, temporal reasoning, and visual change detection over

Artificial Analysis' changelog explicitly logs "New language model evaluation results available — Ling-3.0-flash-VL — Intelligence Index: 25" on 10 September 2026, directly naming inclusionAI's multimodal model and confirming its evaluation date. The changelog also records InclusionAI performance results for Ling-3.0-f This changelog entry provides a reliable, named reference to the subject model on its evaluation date, useful as a primary timestamp and benchmark anchor. It complements other Artificial Analysis surfaces by restating the Intelligence Index 25 figure for Ling-3.0-flash-VL, helping disambiguate the model from text-only