Load Mistral 3 Model

"Load Mistral 3 Model" doesn't generate a single token. It's the loader half of a two-node local-LLM setup, and its whole job is to get a Mistral model onto your machine and hand it to the "Mistral LLM" node. The name is a lie in the good direction: there's no API call, no API key, no cloud - the model lives on your disk and runs on your GPU. That's the entire appeal, and also the catch, because the model it insists on is not small.

What's actually happening

Under the hood this node is three steps in a trench coat. First it checks whether the model already exists in ComfyUI/models/LLM/, and if not it calls snapshot_download from huggingface_hub to pull it from Hugging Face. Notably it passes ignore_patterns="consolidated.safetensors" - the HF repo ships both a raw checkpoint and the transformers-format weights, and this skips the duplicate so you don't download the same ~48GB twice. Then it loads Mistral3ForConditionalGeneration from transformers in bf16 and parks it on ComfyUI's unet offload device (read: CPU RAM) rather than VRAM. Finally it loads the tokenizer from the repo's tekken.json and returns a dict of {model, tokenizer} wrapped in a pack-local M3MODEL type.

That "offload to CPU" bit is worth understanding before you panic about VRAM. The loader is deliberately light on your GPU - the weights sit in system RAM until the LLM node pulls them over for actual generation.

The inputs and outputs (all one of each)

This is the simplest node interface you'll ever meet. There's exactly one input:

• model - an enum with a single hardcoded choice, mistralai/Mistral-Small-3.2-24B-Instruct-2506. There's nothing to configure; the dropdown is furniture for a future multi-model menu. If you're looking at it thinking "surely there's a smaller option," there isn't. Not yet.

And one output:

• mistral3_model (M3MODEL) - wire this into the mistral3_model input of the Mistral LLM node. That's the only thing it connects to.

That 24B model is the same Mistral-3 vision-language family that Flux 2 uses as its text encoder, if you want a sense of what you're getting - a serious, modern LLM, not a toy.

How to install

ComfyUI Manager is the easy path: search "comfyui-mistral-llm" and install. Otherwise:

cd ComfyUI/custom_nodes
git clone https://github.com/otjamin/ComfyUI-Mistral-LLM

Restart ComfyUI. Manager (or your pip install -r requirements.txt) pulls in transformers, mistral-common>=1.6.2, and huggingface_hub. Heads-up: the repo targets Python 3.12 and is young - one commit, an empty README, so treat the code, not the docs, as ground truth.

The part nobody tells you: it's a ~50GB download

Your first run triggers a download of the full bf16 model - roughly 48GB of weights - into ComfyUI/models/LLM/Mistral-Small-3.2-24B-Instruct-2506/. On a decent connection that's a while, so don't judge the node by its first five minutes. snapshot_download is resumable, so an interrupted download will pick back up on the next run rather than restarting.

Where people get burned: the loader happily "works" on a modest card because it keeps weights in CPU RAM - and then the LLM node moves the entire model onto the GPU in one model.to(device) call and dies with a CUDA out-of-memory on anything under ~48GB of VRAM. There's no quantization path in the code, so no Q8/Q4 escape hatch here. If you can't hold the full model, this pack isn't the lever for you - look at a GGUF-based setup instead.

One more thing worth knowing, and it applies to every LLM pack: you're installing code that downloads gigabytes and runs arbitrary Python with your user's permissions. The ComfyUI LLMVISION malware incident in 2024 is why the community is twitchy about this exact category. This repo is small enough to read in a few minutes - that's a habit worth keeping.