The authors investigate whether discrete visual tokens can serve as semantically lossless representations, challenging the prevailing view that quantization error inevitably degrades semantic fidelity compared to continuous visual features. Using multimodal understanding as a proxy for semantic quality, they discover that quantization error alone is an unreliable predictor; some discrete encodings with larger errors outperform tighter quantizations, indicating that the effective bit budget of discrete indices dominates downstream performance. Existing quantization schemes, however, saturate as extra codebooks yield diminishing returns. To address this, they propose **Factorized Vector Quantization (FVQ)**, which splits visual features into complementary, fixed‑dimensional views and quantizes each view with symmetric parallel codebooks. FVQ continues to improve with higher bit budgets and ultimately matches the performance of continuous visual representations in multimodal tasks. Trained on industrial‑scale multimodal data, their model sets new state‑of‑the‑art results among discrete‑token approaches, surpassing **UniAR** by 22.8 % and **Emu3‑Chat** by 59.2 % on a nine‑metric average. Moreover, it is the first discrete‑token model to compete with strong continuous models such as **Qwen3‑VL**, including on fine‑grained OCR benchmarks. The paper outlines the underlying principles and a practical recipe that makes such competitiveness achievable, highlighting the viability of discrete visual tokens for large‑scale native multimodal foundation models. Code and models are publicly released.