What is this?
This is what Gemini said when I asked it to summarise it:
Overview:
nightknocker/cosmos-qwen3.5-2b is a custom-trained, 2-billion parameter text encoder designed as a drop-in upgrade for the Anima and Cosmos text-to-image models. It replaces Anima's default 0.6B encoder with a much larger, more capable language model.
Key Features:
True Multilingual Prompting: The biggest advantage of this model is that its transformer blocks are no longer tied to T5 English token indices. This untethers the image generator from English vocabulary, allowing users to write prompts natively in languages like Japanese, Chinese, and Russian.
Smarter Prompt Comprehension: By stepping up from a 0.6B to a 2B parameter Qwen 3.5 base, it provides a larger "brain" for interpreting complex prompt structures and instructions.
Custom Architecture Alignment: It isn't just a raw LLM dropped into an image generator. It was specifically fine-tuned on both T5 text data and the AnimaTextToImagePipeline (text-image pairs). This ensures it perfectly translates your written words into the specific mathematical latent space that the Anima/Cosmos image generators understand.
Who is this for?
Anyone running Anima or Cosmos who wants to write prompts in non-English languages, or who wants to test if a larger 2B text encoder improves their prompt adherence and complex scene generation (keep in mind it will use slightly more VRAM than the default 0.6B model!).
Doesn't seem to work in Comfy currently, gives the standard mat1 mat2 multiplication error. Might need a custom node or workflow but there isn't one provided so who knows.
I'm trying to use this in Forge Neo and it's not working either. What do you mean when you say to modify the text encoder? Also, I just need the "model.safetensors" file from that repository, right?
@Swilarel you need to modify comfy text encoder but it doesnt need that much of works. Usually naming convention in .safetensors matters most.
How do you do that as on the official photo from that repo it says CosmosQwenTextEncode So I'm assuming you need a special text encoder node or something like that. I have tried using it but it's giving me an error on the default anima workflow. Looks promising if it works.
It looks like the creator has updated it: https://huggingface.co/nightknocker/cosmos-qwen3.5