(Deno) Local LLM Loader

The name is a mild lie: "(Deno) Local LLM Loader" doesn't load a model itself. It calls models that are already running on your PC - Ollama, LM Studio, llama.cpp, vLLM, or any OpenAI-compatible server on localhost - and pipes the answer back into your graph as a result string. No cloud, no API key, no config file to hand the model path to ComfyUI. If you can already chat with a model at 127.0.0.1, this node turns it into a workflow component in about a minute.

Why you'd reach for it

The most common job is prompt work: rewrite or expand a rough prompt, generate negative prompts, summarize, or produce structured JSON that another node eats. The README even points at a concrete pairing - feed JSON prompts from the Local LLM Loader straight into the Ideogram Director. The second job is gating: this node is designed to work with its sibling (Deno) Local LLM Reviewer, where the LLM's verdict decides whether media gets saved.

How it works

Pick provider, then the model. For Ollama and LM Studio, the node hits your local server and lists the actual models it finds, so you're not typing names from memory. For llama.cpp, vLLM, or custom, custom_server_url defaults to http://127.0.0.1:8000/v1 and you name the model in custom_model. There's a hard safety constraint here worth liking: the node only talks to localhost-style hosts (127.0.0.1, localhost), so it won't quietly phone a remote endpoint.

The interesting settings are about memory, not the LLM itself:

• model_memory - Unload after run (default) or keep the model warm. With Keep loaded, keep_minutes controls how long Ollama holds it. For a batch of prompts this matters a lot: keeping the model loaded across the whole batch is dramatically faster than reloading per prompt.
• comfy_vram_policy - whether ComfyUI models get unloaded before the LLM call. Default is Auto: unload only before first LLM call, which is the right starting point on a single GPU.
• thinking - request reasoning from models that support it.
• system_prompt - steer the rewrite/review behavior.
• image - optional, and only useful with a vision-capable local model.

prompt is your main text; the in-node textarea and the STRING socket feed the same backend field, so you can type in the node or wire text in. The output is result - and for batched prompts it returns a STRING list, which is how it keeps one model loaded across many queued runs.

Install and gotchas

Install is just the pack - no extra Python deps, this node is pure HTTP.

cd ComfyUI/custom_nodes
git clone https://github.com/Deno2026/comfyui-deno-custom-nodes.git
# restart ComfyUI, then start Ollama or LM Studio before first use

The number one beginner failure is starting ComfyUI without the LLM server running - the model dropdown comes up empty. Start Ollama/LM Studio first, then hit the refresh button on the node. Second gotcha: not all local servers support the structured-output / reasoning formats the node tries; if you hit a parse error on a particular model, try thinking off or a different model, because that's usually a server capability mismatch, not a broken node. And keep in mind the node sends text only - it never sends AUDIO, even if the rest of the graph has some.