Model query guide: profile facts or unresolved fields
The attached MiniMax H3 Open Weights profile lists these inputs: FL2VA: text with zero, one, or two keyframe images, Ref2VA: text with image, video, and audio references, Repository processor and tokenizer files, and Framework-specific serving configuration. It records these documented capabilities: Local text-to-audio-video, Local first-or-last-frame-to-audio-video, Local reference-to-audio-video, Provider examples for SGLang, vLLM, diffusers, and ComfyUI, and Checkpoint-specific reproducible cases. Its output notes are Released checkpoints use BF16, The repository describes the released checkpoints as CFG-distilled, Local H3-Base is the documented 768p path, and Context-IR and Regenerate-2K require separate hosted modules in the complete workflow. These facts are bounded to MiniMax, the attached references, and their evidence dates; they are not an independent benchmark, a statement about an uncited version, or a promise of access in SEELE, every region, or every account. The profile's bounded workflow use is: Use the open weights when deployment control, reproducibility, or fine-tuning investigation matters and the team can own infrastructure, security, license review, prompt processing, moderation, output review, and the difference from the complete hosted path. Before relying on a listed input, capability, or output, connect it to the particular attached reference whose claim scope covers it and to the same release or service surface. If the reference does not state the fact, or the current interface cannot be reconciled with that source, leave the field Unknown / not verified and keep it out of the test assumption. The profile's repository test briefs are “Provider example reproduction”: Pin a repository revision and reproduce the official task-family example before introducing custom assets. Save environment, framework, checkpoint files, configuration, request, logs, and output hashes., “Local versus hosted boundary”: Run a comparable authorized brief through local H3-Base and the complete hosted path only when input transformation and regeneration stages are documented. Report configurations separately rather than pooling results., and “Operational readiness”: Exercise model loading, request isolation, failure recovery, storage, logging, moderation, rights review, and asset deletion. Treat a successful demo output as only one part of a production deployment decision. Its recorded review checklist is “Repository revision, checkpoint family, framework, precision, and configuration are pinned”, “The current Community License is reviewed for the intended use”, “Context-IR and 2K regeneration are not implied when absent”, “Security, moderation, provenance, storage, and deletion controls are documented”, and “Hardware fit, throughput, cost, and support are measured in the team's own environment”. These are evaluation instructions, not results; this page contains no completed run, output, selection rate, reviewer verdict, or production approval unless one is explicitly attached elsewhere. For “MiniMax H3 open weights,” test only the part of the recorded decision that documentation cannot settle. Use authorized material and preserve the literal input, instructions, visible controls, selected label, interface, account and region, attempt count, failures, output, and observation date. Derive acceptance criteria from “MiniMax H3 local deployment”, “H3 Base checkpoints”, and “MiniMax H3 Community License”, not from a universal score. A result supports only that captured setup; it cannot establish permanent quality, general access, provider affiliation, licensing, safety, or behavior in another version, mode, host, or service.