MiniMax M2.7 Highspeed

MiniMax M2.7 is the 2025 text model that made a name on two things: enormous context and tool calling. The Highspeed variant trades a little ceiling for a lot of latency, which makes it the sensible default for the jobs people actually wire into a ComfyUI graph - prompt tuning, captioning, workflow glue - where you want the answer back before you lose interest. This node runs it on Runware's cloud. No weights, no VRAM, no local LLM server to keep alive.

It's one of several hosted LLM nodes in the pack (this one, Qwen3.5-27B, and a few others), and the honest way to choose between them is: Qwen is the vision pick because it takes IMAGE input; MiniMax is the text-and-tools pick because it leans into long context and structured tool calls. Different jobs, same socket.

The inputs that matter

• messages (required, RUNWARE_MESSAGES): the conversation, built with the Runware Messages builder and chained in.
• settings.maxTokens (default 32,768, up to 196,608): the headline number. That's a 196K-token ceiling - the "paste a whole book in" mode. The default is already generous for prompt work.
• settings.systemPrompt (multiline): set the persona. "You are a workflow assistant that returns JSON only" - that sort of thing.
• settings.temperature (default 1, range 0–1): this one defaults high. For deterministic workflow glue, drop it toward 0; only keep 1 when you want the model's personality.
• settings.topP (default 0.95): the usual nucleus sampling control.
• toolChoice (gated toggle): flip it on to control tool calling. toolChoice.type is the strategy - auto, any, tool (force a specific tool), none - and toolChoice.name names the tool, which the tooltip notes is required when type is tool.
• advanced_json: the escape hatch, and its tooltip is upfront that it's for tools - define your function schemas as JSON here, then control them via toolChoice.
• seed, numberResults (max 4), includeUsage: the usual reproducibility and diagnostics block.

Output is a single text (STRING).

How it works

Pack-standard: messages ride up as a REST request, MiniMax M2.7 Highspeed runs on Runware's GPUs, and the text comes back. includeUsage is the token-meter - flip it on when you're tuning prompts and want to see what the answer cost. And because this is Highspeed, the whole round trip is fast enough to feel interactive rather than like a batch job.

Install and gotchas

Install the pack once - ComfyUI Manager, search Runware - or git clone https://github.com/Runware/ComfyUI-Runware into custom_nodes + pip install -r requirements.txt, restart, and set a Runware API key (Settings → Runware API key, RUNWARE_API_KEY, or runware auth login).

Two things to remember. A 196K-token context is a capability and a bill - pasting a whole novel in and asking for a summary is genuinely useful, and it costs accordingly. And tool calling is a two-part setup: the tools schemas go in advanced_json, and toolChoice tells the model how to use them; forgetting either half makes the other silently pointless. For fast, long-context text work inside a workflow, this is the node to reach for.