Seonglae Cho‣
3.9 2.4T parameters
Qwen 3 VLM NVIDIA Cosmos
Comfy UI-only build that sacrifices text performance entirely while mostly preserving vision performance, enabling extreme ternarization through task specialization. with leaving vision encoder frozen and quantize transformers to ternary bits.
3.8 uncensored
Removed half of the 512 experts from the 176.9B MoE model and quantized the remaining weights to 3.5 bpw. The 354GB BF16 model becomes a 58.4GB file, with a resident working set of about 29.6GB.
flash next gguf
