ModelMergeCosmos14B

ModelMergeCosmos14B is the big sibling of ModelMergeCosmos7B - the block-merge node for NVIDIA's 14B Cosmos diffusion model, with 36 transformer blocks instead of 28. If you're in the small, serious crowd doing Cosmos world-model work, the 14B is the higher-quality of the two variants, and this is the node that lets you blend two of them block-by-block.

Same honest framing as the 7B version applies, and it's worth repeating because it saves you time: Cosmos was built for Physical AI - robotics, simulation, autonomous driving - not for the stable-diffusion-style image community, and it never developed a real merge culture. Nobody on CivitAI is downloading "Cosmos 14B realistic merge #47." You'd use this node because you have a specific technical need: two Cosmos 14B checkpoints, usually fine-tuned for different scenes or conditions, that you want to interpolate between. It's a precision instrument for a niche, and it's entirely okay if it's not your instrument.

How it works

Identical mechanism to the whole family - a ModelMergeBlocks subclass, per-tensor interpolation result = ratio · model1 + (1 − ratio) · model2, ratio chosen by longest matching key prefix. The 14B key map is the same structure as the 7B, just bigger:

• pos_embedder., extra_pos_embedder. - the two positional embedding paths (the second handles Cosmos's temporal dimension).
• x_embedder. - input patch embedding.
• t_embedder. - timestep embedding.
• affline_norm. - normalization layer (upstream's spelling).
• blocks.block0 through blocks.block35 - 36 transformer blocks.
• final_layer. - output projection.

Every slider defaults to 1.0 (all model1). The extra eight blocks versus the 7B are why the two nodes are separate - the shapes don't line up, and merging across sizes fails loudly.

The inputs that matter

Same as the 7B: model1, model2, and a range of blocks.blockN ratios is the 95% case. If you're interpolating between two scene-conditioned fine-tunes, the common play is to hold the early blocks (structure) of one and let the later blocks (detail, texture) follow the other. The extra_pos_embedder. matters if the two parents differ in temporal behavior - for a video/world model that's often exactly what you're trying to blend.

One MODEL out, into your video sampler or Save Checkpoint.

Where people get burned

Same as the 7B, plus a bigger memory bill - you're holding two 14B models plus the patched result, so plan your VRAM. Shape mismatches are the other sharp edge: Cosmos 7B weights do not merge with 14B weights, and Cosmos-Predict2 (a separate, newer NVIDIA family, the one that matters to anime people via Anima) has its own nodes. Pick the node that matches the checkpoint you actually loaded - ComfyUI will happily let you wire a 14B into a 7B merge node and then refuse at sampling with a shape mismatch that looks cryptic.

How you get it

Core ComfyUI, model/merging/model specific, landed February 2025 with Cosmos support. Nothing to install.