Llama.cpp Model Profile

The compact generation nodes in this pack are designed around a single idea: stable settings live behind typed sockets, not scattered across a hundred widgets. Llama.cpp Model Profile is the most important of those sockets, because it's the only required one. It bundles the chat handler and the sampling defaults for a specific model into one connection, and when you pick a profile you get sampling values that were tuned for that model rather than a generic starting point.

The profile combo is where you choose your model family: General (0.2 temperature, top_p 0.95, top_k 40 - the safe default), Gemma 4 Vision and Muse Glimmer (both 1.0 temperature, top_k 64), Qwen 3.5 Thinking (1.0 temp, top_k 20, presence_penalty 1.5, reasoning on), Qwen 3.5 Non-thinking (0.7 temp, top_p 0.8, reasoning off), and Qwen 3 VL (0.7 temp, top_p 0.8). The Qwen profiles carry a presence penalty of 1.5 - that's a genuine choice, pushing the model toward novelty - and the reasoning mode travels with the profile for the Qwen 3.5 pair. When you select Custom, the node reveals the custom_handler and six sampling inputs so you can roll your own; switching back to a named profile keeps your custom values stored without applying them, which is a nice touch if you're experimenting.

Two details that show the care here. First, presence_penalty is forwarded to the targeted JamePeng fork under its API spelling present_penalty - the docs call that out explicitly, so don't be confused if your sampling logs show a slightly different key name. Second, the Qwen 3 VL card doesn't prescribe a min_p, so its profile sets it to 0.0 - no extra minimum-probability filter imposed. The tooltips and docs are unusually precise about this stuff.

Wiring is simple: connect model profile to the model_profile input of either compact Llama.cpp Generate or Sequential Generate. It's required on both, so you'll always see it.

Setup is the pack standard:

cd ComfyUI/custom_nodes
git clone https://github.com/craftingmod/ComfyUI-Ollama-ImageList

Restart ComfyUI afterwards. The usual llama.cpp prerequisites apply: a JamePeng multimodal llama-cpp-python wheel in ComfyUI's Python, and ComfyUI 0.19.3+.

Honest verdict: this node is 90% of the reason to use the compact flow instead of the legacy detailed one. The profiles encode real-world tuning decisions - Qwen 3.5's presence penalty isn't something most people would arrive at by trial and error in an afternoon - and "pick your model from a dropdown, connect it, done" beats scrolling thirty widgets every time. If you're on Gemma 4, the matching Gemma 4 Runtime Preset node pairs with it to handle the batch/token side of that model's quirks. Start with General if you're not sure; it's tuned to be stable, and you can always switch to Custom later.