MiniMax maps H3 from one 24GB card to SGLang, and keeps the interpreter in-house

Continuity is what makes that index worth reading. A team that prototypes on a single 24GB card and then finds its serving stack wants a different model pays for the move twice, once in engineering time and once in output drift. The route MiniMax has cataloged keeps the same weights underneath from a ComfyUI prototype through to SGLang or vLLM-Omni ^[10], so the local step is a rehearsal for the production step rather than a demo that gets thrown away.

The load being rehearsed is heavier than the word "clip" suggests. MiniMax specifies four to 15 seconds at 24 frames per second, 32 kHz stereo, and output to 2K via a regeneration pass ^[7], which puts a maximum-length generation at 360 frames with synchronized audio attached ^[13]. The input side has a trap in it: the reference variant advertises nine image slots, three video and three audio, but the ceiling is 12 files ^[9], so three of those 15 advertised slots can never be filled in the same request ^[14]. A workflow written against the per-type maxima fails at submission, not at generation.

Then there is the part the index cannot route around. MiniMax published H3-Base and supporting inference resources, but H3-Context-IR, the system that works out how a user's text, pictures, audio and reference videos relate to each other, stays hosted; MiniMax's stated reason is that it depends on multiple models and services, and the alternative it offers is to build your own preprocessing from the published prompting guidance ^[1]. So the well-documented local path covers generation and stops at interpretation. On a multi-GPU SGLang deployment, the expensive compute is yours and the semantic front door is still a metered call, unless you write that front door yourself.

The sequence is legible in the dates. The index landed 22 days after the weights ^[15], long enough for independent developers to have produced the awkward parts the navigation now covers: local execution, Apple Silicon, node packaging, acceleration and fine-tuning ^[6]. Read next to Yan Junjie's stated ambition, in MiniMax's 2025 financial-results release, to "evolve from a large-model company into a platform company for the AI era" ^[3], a community-maintained compatibility map is cheap distribution. It costs a README and returns an ecosystem.

What the repository actually proves is breadth of effort, which runtimewire argues is the signal worth having: developers are grinding on the operational details that decide whether open weights are still in use after launch week ^[12]. That is a better test than any adoption figure MiniMax could offer, and it is the one thing the index measures honestly.