Gemini 3.1 Flash Lite

Gemini 3.1 Flash Lite is the "Lite" in Google's Gemini lineup, and in the Runware pack that means one thing: the cheapest text node on the menu. It's the model you wire in when the job is high-volume and low-stakes - batch captioning, tag normalization, prompt cleanup, routing decisions - and you'd rather not pay Pro prices for work a small model can do. The Lite suffix is doing honest work here: it's smaller, faster, and cheaper, and it shows on genuinely hard reasoning.

The honest take: don't give it your hardest task. Where Gemini 3.1 Pro earns its higher price tag on complex multi-step reasoning, Flash Lite shines at the mechanical stuff your workflow does a hundred times - and for that, it's genuinely the right node. Pair it with a clear settings.systemPrompt that pins down the output format and it'll surprise you with how much of the grunt work it handles.

What you set

• messages (required) - the conversation from the Runware Messages builder: role + content pairs, chained for multi-turn.
• images - optional IMAGE input for vision tasks like captioning frames.
• settings.systemPrompt - the format contract. This matters more on Lite than on any bigger sibling; the small model leans on instructions.
• settings.maxTokens - up to 65536, default 4096.
• settings.temperature (0–2, default 1) and settings.topP (0.95).
• settings.thinkingLevel - minimal/low/medium/high. Keep it low for speed.
• toolChoice - gated tool-calling for structured automation.
• includeUsage - token stats in the response, useful when you're cost-tracking.

Output is text, a STRING. advanced_json covers extra inputs (audio/document/video), stop sequences, and tools.

How it works

Standard pack machinery: taskType: textInference over REST through the Runware SDK, reply back as a string. Cost prints on the title bar, which is where Lite's reason for existing becomes obvious - the per-call price is the lowest in the text family. Each call is independent, so state lives in the Messages chain you feed it. One practical pattern: use Lite as a prompt preprocessor - clean up a user's messy input, hand the result to a bigger model or a sampler downstream, and let the cheap node absorb the grunt work.

Installing

ComfyUI Manager → search Runware → install → restart. Manual:

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

No model downloads; deps are runware-sdk, pillow, soundfile. API key via Settings → Runware API key, RUNWARE_API_KEY, or runware auth login.

Troubleshooting

Format drift is the Lite failure mode - it'll occasionally wander off-schema. The fix is a stricter systemPrompt and a shorter maxTokens; give it room to improvise and it will. If answers are flat-out wrong on reasoning-heavy tasks, that's the model's ceiling, not a bug - escalate to Gemini 3.1 Pro. And if you need live web grounding or JSON output, that's Gemini 3.5 Flash's territory.