Meituan LongCat's FVQ Matches Continuous Vision Models

A new paper from a team interning at Meituan LongCat sets out to answer a narrow question with a large implication: whether discrete visual tokens can carry the same information as the continuous features that most multimodal models rely on. The authors say they can.

"We show that quantization error does not determine semantic fidelity," the paper argues, challenging the assumption that compressing continuous visual features into a finite codebook necessarily loses meaning. In one experiment, a discrete representation with an order-of-magnitude larger quantization error still delivered better understanding than a cleaner one.

What actually matters, the authors argue, is the effective bit budget of the discrete code. The problem: existing multi-codebook methods, residual vector quantization (RVQ) and product quantization (PQ), stop improving as codebooks are added, because each new one carries less information than the last. To get past that ceiling, the paper introduces Factorized Vector Quantization, which "quantizes complementary, fixed-dimensional views of the visual feature with symmetric parallel codebooks."

Trained on industrial-scale multimodal data, the resulting model posts relative gains of 22.8% over UniAR and 59.2% over Emu3-Chat on a nine-metric average. The authors claim it is "the first discrete-representation model competitive with strong continuous-representation models such as Qwen3-VL," including on fine-grained OCR tasks.

Corresponding author Xinggang Wang is at Huazhong University of Science and Technology; the co-authors interned from Peking, Shanghai Jiao Tong, Tsinghua, CUHK and MBZUAI. It arrives in a dense stretch of multimodal research we have been tracking.