MOSS-Image's ComfyUI bundle adds a small VAE and BF16 text encoder
Comfy-Org's downloadable Base/SFT package has a documented folder layout and multiple precision options, with actual October 17-18 uploads behind the latest attention.
ComfyUI users have another concrete route to MOSS-Image: Comfy-Org’s component bundle now includes a small FLUX.2 VAE and a BF16 Qwen3.5 9B text encoder alongside its existing image-model files. The additions are dated October 17 and 18 in the public upload history, rather than inferred from a fresh collection timestamp.
The repository itself began October 16. This is packaging of the existing MOSS-Image Base and SFT models, not a new OpenMOSS foundation-model announcement. It also differs from the separate community distillations and Apple MLX ports: the documented destination here is ComfyUI’s normal model folders.
Choose components, not just one checkpoint
The English card lists Base and SFT diffusion weights in BF16 and int8 ConvRot variants. They go in ComfyUI/models/diffusion_models/. Qwen3.5 9B text encoders go in models/text_encoders/, with BF16, int8 ConvRot and W4A8 GPTQ options. The full and small FLUX.2 VAE files go in models/vae/.
That separation matters when comparing downloads. A W4A8 text encoder does not mean that the entire image pipeline runs at four-bit precision, and one diffusion checkpoint’s file size is not the complete application’s memory requirement. Base and SFT are separately named model choices, not interchangeable labels for a single download.
The history records flux2-vae_small.safetensors being uploaded on October 17 at 18:48 UTC. The BF16 text encoder followed on October 18 at 10:49 UTC. The most recent October 18 entry is a README edit; it should not be substituted for either file’s original availability date.
What this package establishes
The card declares Apache-2.0 and supplies a concise installation layout. It does not provide a minimum supported ComfyUI version, a complete node graph or a hardware-specific performance table. Users still need a compatible workflow and sufficient resources for all of its components.
AITrending has not run this package. The filenames and folder instructions support a useful local-image deployment story, but do not establish that the smaller VAE is faster, that int8 preserves every detail, or that a particular GPU can load the full pipeline. The news is the inspectable component selection and dated weekend additions, without importing performance claims from unrelated MOSS-Image conversions.
