Kimi K2.6

Kimi is Moonshot AI's model line, and the K2 series has a reputation for strong open-weights reasoning - the kind of model people use for the thinking-heavy jobs that small local LLMs embarrass themselves on. K2.6 is the current tier, and this node puts it inside ComfyUI as a text node with a serious spec: an up-to-1,000,000-token output ceiling via settings.maxTokens_value, five levels of settings.thinkingLevel, tool calling, JSON schema output, and vision via an IMAGE input.

Why would a ComfyUI user want this? Same reason you'd want any good LLM node: prompt rewriting, captioning, structured JSON generation for downstream nodes, and decision logic in multi-step workflows. K2.6's edge is the combination of a huge context, cheap reasoning levels, and JSON output - which makes it genuinely useful as a workflow brain that hands structured data to the rest of the graph.

How it works

The standard pack text-node path: required messages socket (Runware Messages builder: role + content, stackable), a textInference request, and a STRING text output. Where K2.6 stands out is the control surface - it exposes more thinking and output control than the other text nodes in the pack.

Inputs that matter

• messages (required) + settings.systemPrompt - conversation and system instruction. For structured-output workflows, put the format spec in the system prompt and use outputFormat: JSON.
• settings.thinkingLevel - none, low, medium, high, xhigh (default none). This is the knob that makes K2.6 interesting: it's a reasoning model, but unlike some peers it lets you dial reasoning from off to extreme. For pipeline tasks, none/low is fast and cheap; for hard analysis, high/xhigh. Defaulting to none is a nice "don't pay for thinking you didn't ask for" default.
• outputFormat - TEXT or JSON. JSON mode is the underrated one: combined with the advanced_json jsonSchema field, you can get reliably-structured output out of a text node - hugely useful when the result feeds other nodes.
• settings.maxTokens - off-by-default gate with maxTokens_value (up to 1,000,000). That ceiling is enormous; leave the gate off for the model default, and only flip it on if you're doing something genuinely long. One careless 1M-token call is a bill, not a result.
• images - an IMAGE input, so K2.6 is vision-capable: wire a rendered image in and ask for analysis, captioning, or critique.
• settings.promptCacheKey - a cache key for reusing prompt processing across requests. If you're running the same big system prompt repeatedly (batch captioning, same-context loop), set a key and the repeated prefix gets cached - a real cost saver for cloud LLM use.
• toolChoice / toolChoice.type / toolChoice.name - tool calling via advanced_json (tools). auto/any/tool/none.
• settings.temperature (default 1), settings.topK (default 0), settings.topP (default 1), minP, the three penalties - full sampler set. For deterministic pipeline work, drop temperature to ~0.2.
• numberResults (1–4), seed, includeUsage - variations and spend visibility.

Single output: text (a STRING - if you chose JSON mode, it's a JSON string; parse it in a downstream node).

Installing

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

Or ComfyUI Manager → search Runware → install → restart. API key via Settings, RUNWARE_API_KEY, or runware auth login.

The honest caveats

The 1M-token output ceiling is real but you will almost never touch it - and the cost scales with what you use, so default to small. The advanced_json fields (jsonSchema, tools, settings.stopSequences) are where half the power lives, and they're easy to miss if you only look at the visible widgets. And remember the output is a plain string: wiring K2.6's answer into a workflow that needs numbers or booleans means you're parsing text (or using JSON mode) yourself. For "big-brain reasoning that hands structured answers to my graph," this is one of the best values in the pack.