🤖 Nexa Text Generation

This is where the pack's Nexa SDK mode actually talks. NexaModelSelector finds the models; this node generates the text. Wire the selector's model_config output into this node's model_config input, pick a model, type a prompt, and you get two strings back: context (the answer) and thinking (the reasoning, if you asked for it). That split is the best thing this node does - reasoning models like DeepSeek-R1 and the Qwen3-Thinking line vomit their chain-of-thought into the output, and here you get the clean answer without the noise, or with it, on a separate wire, whichever you want.

How it works

The node calls the Nexa service's OpenAI-compatible /v1/chat/completions endpoint - no key, no cloud, just your local service on port 11434. The preset_model dropdown is built at load time: downloaded models at the top, then a separator, then the pack's preset model IDs. Those presets are all Qwen3, mostly abliterated variants, mostly Q8_0 - the pack's taste, so pick accordingly. Select Custom and type your own ID into custom_model in author/model:quant format. auto_download (on by default) runs nexa pull for you when the model isn't cached.

Then comes the clever bit: the node strips thinking tags - <think>…</think>, <thinking>…</thinking>, [THINKING]…[/THINKING] - out of the response. context gets the cleaned final answer; thinking gets whatever reasoning was stripped, and only when enable_thinking is on. Set it off and the tags still get removed, you just don't get a thinking wire filled.

The inputs that matter

• model_config - from NexaModelSelector; without it, nothing downstream works.
• preset_model / custom_model - one or the other; presets are a dropdown, custom takes an ID.
• auto_download - keep on unless you deliberately pre-pull everything.
• max_tokens - default 512, but the README makes a strong case for 256: at 512 you get eleven paragraphs of boilerplate, at 256 you get one clean paragraph. The whole "why 256" section is worth a read.
• temperature (0.7), top_p (0.9), top_k (40), repetition_penalty (1.1) - defaults are fine to start.
• enable_thinking - on if your model reasons and you want it separate.
• prompt - at the bottom of the node, deliberately, because it's the field you edit most.

Common issues

"Service is not available" means nexa serve isn't running - start it and confirm with curl http://127.0.0.1:11434/v1/models. Output running long? Cut max_tokens to 256 and tighten the system prompt. A model not in the dropdown? nexa pull it, refresh the selector, restart ComfyUI. If nexa list shows 0B entries, it's the stale-cache cleanup from the README (~/.cache/nexa.ai/nexa_sdk/models - nuke the local/workspace dirs and the *.lock files).

Install

ComfyUI Manager (search "ComfyUI-GGUF-FX") or clone + pip install -r requirements.txt, restart. Nexa mode additionally needs pip install nexaai and a running nexa serve - that's the price of admission for the managed-model workflow. If you just want to run a GGUF file directly without a service, the pack's local text path is the lighter option.