MiniMax H3 CLIP (CRT AutoDL)

MiniMax H3 doesn't encode text with a cute little CLIP - it uses a Qwen3-VL 32B text encoder, and that thing is a giant. MiniMax H3 CLIP (CRT AutoDL) is the node that loads it for you, in two quantized flavors, with the auto-download handling the heavy lift.

This is a "get it from the right place" node, not a magic one: it's one of PGCRT's CRT-Nodes AutoDL family, so on first run it fetches the checkpoint into ComfyUI/models/text_encoders/ and loads it as a CLIP. You still need to wire it into the H3 stack - model from MiniMax H3 Model, video and audio VAEs from the sibling nodes - but the file-hunting is done.

The one decision: which quant

model_name has exactly two options:

• INT8 (default) - qwen3vl_32b_minimax_h3_int8_convrot.safetensors, from Comfy-Org's MiniMax-H3 repack. This is the standard choice for ComfyUI's H3 support.
• NVFP4 - qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors, an AWQ-tuned 4-bit version. This is the memory-lean option, and the pairing people use when they're squeezing H3 onto a smaller card.

The encoder is often the surprising VRAM hog in H3 workflows - 32B of text encoder can rival the DiT. If you're on a 16 GB card and picking, the NVFP4 encoder plus a Light W4A8 model is the combination that actually fits. On 24 GB+ with room to spare, INT8 is the safe default.

Output is a single CLIP, which feeds the CLIP Text Encode side of your H3 graph. It doesn't encode anything itself - like all CRT AutoDL nodes, it's a loader with one job.

Installing it

Part of CRT-Nodes - install once:

cd ComfyUI/custom_nodes
git clone https://github.com/PGCRT/CRT-Nodes.git
pip install -r requirements.txt

or via ComfyUI Manager (search CRT-Nodes), then restart. The pack's requirements are heavy, but nothing here is CLIP-specific. The real wait is the first-run download - a 32B encoder in INT8 is still a multi-GB file, NVFP4 smaller but not small.

Common issues

• "Which one do I pair?" - INT8 CLIP with the standard INT8 model; NVFP4 CLIP with a Light W4A8 model (or an NVFP4 setup) for the lean combo. Mixing NVFP4 text encoding with a full-precision model is legal but pointless.
• First run hangs - it's the download; the console shows progress.
• License reminder - the H3 weights carry a community license that excludes the US, EU, UK, and Korea from the applicable territory. The CLIP rides along with the rest of the H3 stack, so that restriction applies here too.
• Update breakage - red node or NaN → right-click → Fix node (recreate).