LM Studio Configuration

This node does nothing glamorous, and that's exactly the point. LM Studio Configuration is the shared settings card the rest of the MiniMax H3 Novel pack hangs off: Extract Chapter References, Consolidate References, and Generate H3 Prompts all take its single lmstudio_config output as an input. You drop one in, wire it to everything, and never touch per-node server settings again. If you've installed this pack and your first run errors out before doing anything, the cause is almost always sitting right here.

The pack doesn't run diffusion at all - it runs a local LLM (Qwen) through LM Studio's OpenAI-compatible server. That whole chain of calls is what reads your novel text and eventually writes MiniMax H3 prompts. So the config is the connector to that brain, not to a checkpoint. api_url defaults to http://127.0.0.1:1234/v1, which is where LM Studio's local server listens out of the box. If you run LM Studio on another machine or changed the port, this is the one field you edit.

The part people trip on: the API key

The key is deliberately not a widget on this node. It lives in ComfyUI Settings → MiniMax H3 Novel → LM Studio → API Key, gets pushed to the server once, and is held in memory only - never written into the workflow JSON. That's a genuinely thoughtful choice; a surprising number of local-LLM and API-wrapper nodes bake credentials straight into the graph, which is a nasty habit in a scene that has already been burned by malicious key-hungry custom nodes. LM Studio accepts the literal string lm-studio as its default key, so if you never configured one, set that and move on. Forget it and every downstream node dies with a French error - the author is francophone, and so are the tooltips here, which is also why you'll see odd accents in the UI. It's not a bug.

The Qwen3.5 knobs, briefly

Everything past api_url and the optional model field exists because local Qwen GGUF models in LM Studio are genuinely fiddly: some LM Studio builds ignore enable_thinking=false, and reasoning streams mangle structured JSON. The defaults are the safe ones - thinking off (faster, and JSON extraction doesn't need hidden reasoning), chat_backend on auto, which picks a transport that works around the thinking-toggle problem for Qwen3.5-style models. You'll normally only touch these if output comes back truncated: qwen35_length_retries (2) reruns with a compact-output instruction, qwen35_max_output_tokens (3500) is the safety cap, and qwen35_safe_chunk_chars (3600) keeps prompts short enough that Qwen stops before running out of tokens instead of mid-JSON. Leave qwen35_min_p at 0 and qwen35_top_k at 20 unless you're chasing a specific failure.

The optional model string is the exact model ID loaded in LM Studio. Blank means auto-select - the first loaded model - which works until you have several loaded and it picks the wrong one. Set it explicitly if the results ever look like they came from a different brain.

Outputs and what to do with them

You get lmstudio_config (type MINIMAX_LMSTUDIO_CONFIG) and a configuration_status string that summarizes what got assembled - a quick sanity check that the URL, model and backend are what you think they are. Wire the config object into the three pipeline nodes. That's it. No images, no latents, no model weights to download for this node - the heavy lift is the Qwen model you load inside LM Studio itself, which the pack never downloads for you.

Common failure modes, in order of likelihood: the key isn't set (see above); LM Studio's server isn't running or the model isn't loaded; the model ID in model doesn't match what LM Studio reports; and you left thinking on and everything got slow and rambly. Start there and this node becomes a set-and-forget.