ByteDance Seed משחררת את Cola DLM: מודל diffusion לא-אוטורגרסיבי של 2B פרמטרים

Cola DLM is architecturally unusual. Instead of predicting one token at a time conditioned on prior tokens (the autoregressive recipe behind nearly every production LLM), Cola plans a passage in continuous latent space using a diffusion process and then emits tokens in a single decode step. The claim is meaningfully faster generation and structurally better long-form coherence, because the model commits to a passage-level plan before realizing surface text.

The 2B scale keeps it firmly in research-prototype territory — no one is putting Cola against GPT-5 — but it's the first openly released non-autoregressive language model with a workable recipe and code. For HF/local-LLM researchers, that's the more important fact: a reproducible alternative to the autoregressive monoculture, with weights available to fine-tune.

Elsewhere on Hugging Face this week, Tencent ARC Lab released Pixal3D (May 11), a model that generates high-fidelity 3D assets from a single image via pixel-feature back-projection, with training code, data toolkit and a Trellis.2-based improved version. Together, Cola and Pixal3D suggest the Chinese AI research ecosystem is using HF aggressively as a distribution channel for architecturally distinctive open releases — even as Alibaba's Qwen frontier tier goes paid-API.