MiniMax H3 RAVEN Streaming Profile / 流式参数配置 (Advanced EXP/T8)

This node is the parameter hub of the RAVEN streaming chain: it emits one synchronized set of sampling parameters that feeds both the T8 audit node and the external RAVEN Streaming Sampler, so you tune in exactly one place instead of chasing the same five numbers across three nodes. Its whole job is to be the single source of truth - and, by default, to keep you on the published profile instead of free-styling.

How it works

preset defaults to published_preview_4nfe, which fixes the reviewed configuration: 4 steps, video/audio shifts 12/3, sink/window 2/2, CPU-pinned KV. That's the profile the README validates as the T8-unified 4-NFE route. The alternative is manual_experimental, which forwards whatever you type in the manual_* fields exactly as-is - and the tooltip is careful to call that unreviewed. The node won't stop you, but it won't pretend you're on the blessed path either.

Outputs mirror the inputs so you can fan them out: steps, video_shift, audio_shift, sink, window, kv_cache_storage as typed values, plus a profile_json string for the audit. Wire the typed outputs into the RavenRequestAudit node and the same values into the external sampler, and the two can't drift.

Why the default matters

RAVEN's published preview is a 4-NFE flow-matching setup - the whole "few steps" trick that makes streaming H3 practical. Crank manual_steps to 20 because that's what you're used to from stock sampling and you've left the reviewed envelope. There's a version of that that works, but it's not this profile, and the pack's philosophy is to make the unreviewed path visible rather than seamless. The shifts deserve special attention too: video_shift 12 / audio_shift 3 are the H3 dual-clock defaults and the sane place to start; if you change one, change the audio one to match what your model actually wants.

Installing

In the T8mars MiniMax H3 Audio T8 pack:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8

or ComfyUI Manager → "MiniMax H3 Audio T8". No pip deps. It's a pure parameter node - no model loading, no VRAM cost, no GPU time. That makes it the cheapest node in the RAVEN stack to tinker with, which is exactly why it exists: experiment here, not by editing a sampler mid-run.

The take

Honestly, the value is discipline, not magic. On the default preset this node is a glorified constant; its real worth is the moment you decide to experiment, because it forces you to acknowledge "manual_experimental" and keep the audit and sampler on the same numbers. If you're not using RAVEN at all, skip this entirely - it does nothing in a stock H3 workflow.