Ollama Generate (deprecated)

This is the node the pack shipped with before it got split into the composable Connectivity/Options/Generate family. If you're seeing OllamaGenerate in a workflow someone shared, especially anything from around 2024, this is what they used - it's a single node with the server URL, model picker, and generation settings all baked directly into it. No chaining, no separate options bundle, just a prompt in and a response out.

It's marked deprecated now. The pack author's own guidance is blunt about it: "old V1 nodes are still available, please replace them with the above ones" - meaning OllamaGenerateV2. It still works, but it's frozen where it is.

Why it got replaced

Putting the server URL and model directly on every generate node was fine for a one-node workflow, but the moment you wanted two or three LLM calls chained together - say, one node drafting a prompt and another critiquing it - you had to re-enter the same server address on each one, with no way to share settings between them. OllamaConnectivityV2 and OllamaOptionsV2 exist specifically to stop that duplication: set the server once, reuse it everywhere.

Inputs and outputs that matter

• prompt (multiline, default "What is Art?") - your only real prompt control. Notably, there's no separate system-prompt field on this node at all; if you want to steer behavior beyond the prompt itself, you have to fold instructions into it directly.
• model - a dropdown populated from your Ollama server's installed models.
• url - defaults to http://127.0.0.1:11434.
• keep_alive (-1 to 60) - how long, in minutes, the model stays loaded after this call.
• format - text, json, or unset, for structured output when you're feeding the response into further processing.
• filter_thinking (default true) - strips a reasoning model's <think> scratchpad out of the output automatically, so response comes back clean instead of mixed with the model's internal reasoning.
• debug - enable/disable console logging.
• Output: a single response string.

Installing it

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

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

then restart. Separately, you need Ollama itself running somewhere reachable - install via the curl script (Linux), the desktop app (Mac/Windows), or Docker - with at least one model pulled onto it.

Common issues

"Value not in list: model." This is the single most commonly reported failure across this pack's older nodes, and it's mechanical: the model dropdown gets baked into the workflow JSON at save time. Load a shared workflow, and the model name saved in it almost never matches what's actually installed on your server. One real thread on this exact problem shows someone hitting it three ways at once - a stale model name, a temperature value outside the allowed range, and an empty keep_alive value - all from reusing a workflow built on a different setup. The fix that actually worked for people: delete the node and re-add it fresh from the menu so it pulls a current model list, then match your models to what the workflow expects (or just retype the prompt on a fresh node).

Reasoning text leaking into the response. If you're on a <think>-capable model and getting a wall of internal reasoning mixed into your output, check that filter_thinking is actually on - someone may have shared a workflow with it disabled.

No further development. This node isn't getting a system prompt field, vision support, or chained history - those all live in OllamaGenerateV2 and OllamaChat now. If you outgrow the basics here, that's the direction to move, not a setting to dig for on this node.