GLM-5.3
Z.ai (Zhipu AI)GLM-5.3 uses the same base model as GLM-5.2 — every gain comes from post-training.
general family · filed under post-training
Post-training is a second training stage that follows discrete-token pre-training.1 It specializes the model to the target use-case while adding a flow-matching action expert for continuous actions, training jointly with next-token prediction to keep text capabilities.2
Trade-off Jointly training with next-token prediction preserves the model's text prediction capabilities.3
Summarised by glm-5.3-flash from 1 source; every sentence is checked against the quote it cites.
After pre-training the model with discrete tokens for 280k gradient steps, we perform a second stage of training that we refer to as post-training— π0.5\pi_{0.5}: a Vision-Language-Action Model with Open-World Generalization
to both specialize the model to our use-case (mobile manipulation in homes), and to add an action expert that can produce continuous action chunks via flow matching— π0.5\pi_{0.5}: a Vision-Language-Action Model with Open-World Generalization
This stage jointly trains with next-token prediction, to preserve text prediction capabilities— π0.5\pi_{0.5}: a Vision-Language-Action Model with Open-World Generalization
Also called Flow-matching action-expert post-training, hybrid training procedure.
4 quotes from 3 sources, one card per model, each opening with its strongest.
GLM-5.3 uses the same base model as GLM-5.2 — every gain comes from post-training.
Post-training then attaches the continuous action expert to this already action-aware backbone
uses a hybrid training procedure, with discrete tokenized training in the pre-training phase, and training with a flow matching action expert only in the post-training phase
training with a flow matching action expert only in the post-training phase