SmolLM3 Simple Generate

The default prompt on this node is literally Gravity is. That tells you everything about its personality. SmolLM3 Simple Generate is the pack's no-frills text-completion node: no chat template, no temperature, no top_p, no seed. Load the model, type a fragment, and the model just keeps going from where you stopped.

When you'd reach for it

Use it for auto-complete-style work - finish a sentence, continue a story, list ideas from a half-written line, or give it the start of a haiku. It's also the node to pick when you want the same output every single run: because it has no sampling knobs and no seed, it decodes greedily, which means identical input gives you identical output. No randomness to fight.

If you actually want assistant-style answers with a system prompt and sane randomness, skip this and use the SmolLM3 Sampler (Chat) instead. The Simple Generate is the blunt tool; the Sampler is the sharp one.

How it works

Mechanically it's the thinnest node in the pack: tokenize your prompt, call model.generate with only max_new_tokens, decode, return. Because do_sample is never enabled, generation is greedy - the model picks the single most likely next token at every step. That's deterministic, but it's also why long generations tend to drift into boring, repetitive loops. Keep your prompts short and your token budget honest.

The only real input besides the model pair is max_new_tokens - default 128, and this node caps at 2048 (the Sampler goes up to 32768). That cap is fine; on a 3B model, greedy decoding past a few hundred tokens usually isn't worth reading anyway.

The output gotcha

generated_text here is not the same thing as the Sampler's generated_text. The Sampler strips your prompt and hands you just the new tokens. Simple Generate decodes the entire sequence - so the output string includes whatever you typed as a prompt, followed by the completion. That's by design, and it's actually useful if you want the full finished sentence in one string. Just don't be confused when the answer seems to quote you back at yourself.

Install

Same as the rest of the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/XmYx/ComfyUI-SmolLM3.git
# restart ComfyUI

Or search "SmolLM3" in ComfyUI Manager, if it's in the registry. And same prerequisites: first run downloads the ~6 GB SmolLM3-3B weights from HuggingFace, and you want a reasonably current transformers (the repo pins nothing, so a stale one from another node can break the load).

Bottom line

This is the node for when you want a text model to complete, not converse. It's honest about being minimal - which is exactly its appeal if your use case is "finish this fragment deterministically" and you don't want three sampling dials in the way.