GLM-4.6V-Flash Modified Tokenizer

This is a modified version of the tokenizer for GLM-4.6V-Flash. You can find the full details for the parent model here.

What's Changed

The chat template has been modified to preserve reasoning content (<think> blocks) from the previous assistant turn in addition to the current one. This creates a "rolling window" of visible thinking, allowing important thoughts to be carried forward across turns without flooding the context with tokens from all previous reasoning blocks.