ByteDance's Seed team has open-sourced Cola DLM, a continuous latent diffusion language model, aiming to innovate text generation by organizing high-level semantics before refining them into specific words. This model, comprising 2.3 billion parameters, utilizes a Text VAE and a block-causal DiT to map text into a continuous latent space and learn latent priors, respectively. Cola DLM has shown competitive performance across eight benchmarks, including LAMBADA and SQuAD, but remains a research checkpoint without instruction fine-tuning or RLHF. The model's diffusion process focuses on latent semantic representations rather than token-level denoising, marking a departure from traditional left-to-right generation paths. While the open-source release includes only the text pipeline, preliminary experiments suggest potential extensions to unified text-image modeling. Cola DLM's primary purpose is to explore the application of continuous latent diffusion in text generation.