Ollama Options

If OllamaConnectivityV2 is the checkpoint loader for your LLM setup, this is the sampler node - the bundle of generation knobs (temperature, top_k, top_p, mirostat, and a dozen more) that map directly onto Ollama's own API options. Plug its output into the options input on OllamaGenerateV2 or OllamaChat and every option you've turned on rides along with the request. The pack's own docs point straight at Ollama's Modelfile parameter table for what each one does - this node is just a ComfyUI-shaped front end for that same list, plus one extra field that isn't part of Ollama at all.

How it works

Every option here comes as a pair: the value itself, and an enable_<name> boolean next to it. If the enable toggle is off, the value is ignored completely - Ollama falls back to its own defaults regardless of what the slider says. This is the single most important thing to understand about this node, because it's also the most common way to get confused by it: cranking temperature up and seeing no change in the output almost always means enable_temperature was never flipped on.

The options a beginner will actually touch

• temperature (default 0.8) - higher makes the model more creative/random, lower makes it more deterministic.
• top_k (default 40) and top_p (default 0.9) - work together to control how wide a net the model casts when picking the next token. Higher values, more variety; lower, more focused and repetitive-safe.
• num_predict (default -1, meaning unlimited) - caps how many tokens the model can generate. Set this if a model tends to ramble past where you want it to stop.
• num_ctx (default 2048) - the context window: how much of your system prompt + conversation history the model can actually see at once. If responses seem to "forget" earlier instructions, this is usually why - raise it, at the cost of more memory on the Ollama side.
• seed - fix this for reproducible output on the same prompt.
• repeat_penalty (default 1.1) and repeat_last_n (default 64) - how hard the model is penalized for repeating itself, and how far back it looks to check.
• stop - a string that, once generated, halts output immediately.

Beyond those, you've got mirostat / mirostat_eta / mirostat_tau (an alternative sampling algorithm to top_k/top_p that actively targets a text-quality range - leave disabled unless you specifically want it), min_p (a newer alternative to top_p), and tfs_z. All of them follow the same enable-to-activate pattern.

One field that isn't an Ollama option at all: debug. It's the pack author's own addition, purely for extra console output from the custom node - it does nothing to the actual API call.

Output: a single options socket (type OLLAMA_OPTIONS).

Installing it

Via ComfyUI Manager, search comfyui-ollama (by Stav Sapir) or ollama → stavsap. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/stavsap/comfyui-ollama
pip install -r requirements.txt

then restart. This node doesn't talk to Ollama itself - it just builds a settings bundle - so there's nothing extra to install for it specifically beyond the pack and a running Ollama server for the Generate/Chat node downstream to actually use it against.

Common issues

"I changed a setting and nothing happened." Check the matching enable_ toggle. Every value on this node is inert until its toggle is on.

Responses feel like they've forgotten the system prompt. Enable num_ctx and raise it. The default 2048 is small enough that a long system prompt plus a growing conversation history can genuinely run out of room, and the model just won't see the earlier parts anymore.

Stacking mirostat with top_k/top_p. These are two different sampling strategies, not complementary settings - turning on mirostat and leaving top_k/top_p enabled isn't wrong exactly, but it's not a combination with a clear, well-understood effect either. If you don't have a specific reason to use mirostat, leave it off and stick to temperature/top_k/top_p, which is what most models are tuned around.