Ling-3.0-flash-VL brings images and video to InclusionAI’s sparse model family

Ant Group’s InclusionAI lab has released Ling-3.0-flash-VL, an open-weight multimodal model that accepts text, images, and video within the same context. The MIT-licensed release includes BF16 and FP8 weights, with FP4 and INT4 quantizations planned.

Ling-3.0-flash-VL combines a 124-billion-parameter mixture-of-experts architecture with 5.5 billion active parameters per token and a context window of up to 256K tokens. The active parameter count reduces per-token computation. Local deployment still requires enough aggregate memory to hold the full model weights, runtime overhead, and KV cache.

A 124B model with a 5.5B path

Specification	Details
Inputs	Text, images, and video
Parameters	124B total, 5.5B active per token
Backbone	42 layers with a 5:1 ratio of KDA to Gated MLA
Context	Up to 256K tokens
Available weights	BF16 and FP8
Planned weights	FP4 and INT4
License	MIT

The architecture extends the text-only Ling-3.0-flash backbone with a Vision Transformer encoder and a two-layer MLP projector. The encoder extracts features from images and video, and the projector maps those features into the representation space used by the language model.

VideoRoPE encodes spatial position and temporal order, allowing the model to track where an object appears and how a scene changes across frames. The effective number of images or video frames that fit in a request depends on preprocessing, frame sampling, visual tokenization, and the text already occupying the context window.