MiniMax H3

Start with the complete system definition

MiniMax published H3 on July 31, 2026 and describes a three-part system: H3-Context-IR interprets free-form multimodal instructions, H3-Base generates a 768p audio-video result, and H3-Regenerate-2K uses the original context with the base result to produce a higher-resolution version. That distinction matters because the hosted system and the downloadable base checkpoints are not identical products. Record which path, checkpoint, API, region, and date your evaluation actually uses before comparing an output.

• Provider: MiniMax
• Official release date: July 31, 2026
• Evidence checked: August 15, 2026

Verify the documented output envelope

The official repository lists output durations from four to fifteen seconds, 24 frames per second, a range of landscape, square, and portrait aspect ratios, and 32 kHz stereo audio. It says the shorter side defaults to 768 pixels and associates 2K generation with H3-Regenerate-2K. Treat these as documented system specifications, not a promise that every interface exposes every combination or that a selected result will satisfy a particular delivery standard.

Match the test to the input mode

Use H3-Base-FL2VA for text, an optional first frame, an optional last frame, or both. Use H3-Base-Ref2VA when the brief depends on several image, video, or audio references. A useful evaluation holds the creative goal stable while changing only the input mode, then records identity continuity, requested motion, camera logic, audio timing, text legibility, failure cases, and the amount of editorial repair required.

Keep provider statements and observations separate

MiniMax lists advertising, branding, ecommerce, product design, interface design, and games among H3 scenarios. Those examples establish intended scope, not measured suitability for your campaign or pipeline. Save the exact prompt, authorized references, settings, output, reviewer, date, and acceptance rubric. Describe what happened in that run without turning a small sample into a universal quality, speed, cost, safety, or rights conclusion.

Model query guide: read “MiniMax H3” precisely

The exact repository query is “MiniMax H3.” Its declared reader intent is evaluate; its taxonomy job is model orientation and fit review. No keyword-source attribution is attached to this retained manual entry; its visible page fields and references, when present, are the complete repository record used here. It combines a short possible entity label with one or more qualifiers; those qualifiers describe the reader's question, not documented product properties. The material qualifiers detected here are identity and workflow-fit wording. The supporting topics recorded for this entry are “MiniMax H3 model”, “H3 video generation”, and “omni-modal video model”. Those fields describe why the page exists and what the reader is asking; they are not product documentation. The attached repository profile names “MiniMax H3” and provider “MiniMax,” with status “Official release and repository verified.” The wording overlaps the attached official label, but facts still remain limited to the cited release and date. Keep the literal query, any possible entity name, every qualifier, and the desired decision separate until the provider-controlled identity record supports joining them.

Model query guide: evidence available for “MiniMax H3”

Evidence class: source-backed. Evidence boundary — checked August 15, 2026: model identity, architecture, input modes, and stated specifications come from MiniMax first-party materials. No independent benchmark, SEELE integration record, price verification, customer result, or production-quality claim is supplied. This is an editorial model snapshot, not a live endpoint or an assertion that MiniMax H3 is available through SEELE. 2 dated references are attached from MiniMax Research and MiniMax. Their presence supports only their stated claim scopes. “MiniMax H3: Breaking the Boundaries of Tasks and Modalities” by MiniMax Research, checked 2026-08-15, supports only this scope: Official release date, omni-modal positioning, stereo audio-video generation, stated output ceiling, intended scenarios, and system modules. “MiniMax H3” by MiniMax, checked 2026-08-15, supports only this scope: Official input modes, output specifications, architecture, checkpoints, module boundaries, recommended workflows, and license link. The attached profile states: MiniMax describes H3 as a general-purpose omni-modal generation system that interprets text, images, video, and audio together and produces video with native stereo audio. Its complete documented path separates context interpretation, base generation, and 2K regeneration. Its separate boundary is: This snapshot binds specifications to MiniMax first-party materials checked August 15, 2026. It does not establish SEELE availability, independent quality, current pricing, generation speed, rights clearance, or production suitability. The recorded evidence dates are 2026-08-15. They identify when the attached material was checked and do not make a name, plan, interface, or specification permanently current. Treat every reference by its written claim scope. Identity evidence cannot establish quality; an input document cannot establish price; a provider page cannot establish SEELE access; and an observation from one account cannot establish behavior in another region, version, mode, or host. If a source and the current product surface conflict, retain both records and leave the disputed field unresolved.

Model query guide: repository sections, points, and FAQs

“MiniMax H3 model” remains an editorial question until a claim-scoped source or authorized observation supplies an answer. “H3 video generation” should be converted into a specific input, control, observable output, rejection condition, and dated result. “omni-modal video model” should be converted into a specific input, control, observable output, rejection condition, and dated result. The page organizes the query through The repository section “Start with the complete system definition” says: MiniMax published H3 on July 31, 2026 and describes a three-part system: H3-Context-IR interprets free-form multimodal instructions, H3-Base generates a 768p audio-video result, and H3-Regenerate-2K uses the original context with the base result to produce a higher-resolution version. That distinction matters because the hosted system and the downloadable base checkpoints are not identical products. Record which path, checkpoint, API, region, and date your evaluation actually uses before comparing an output. Its recorded points are “Provider: MiniMax”, “Official release date: July 31, 2026”, and “Evidence checked: August 15, 2026”., The repository section “Verify the documented output envelope” says: The official repository lists output durations from four to fifteen seconds, 24 frames per second, a range of landscape, square, and portrait aspect ratios, and 32 kHz stereo audio. It says the shorter side defaults to 768 pixels and associates 2K generation with H3-Regenerate-2K. Treat these as documented system specifications, not a promise that every interface exposes every combination or that a selected result will satisfy a particular delivery standard., The repository section “Match the test to the input mode” says: Use H3-Base-FL2VA for text, an optional first frame, an optional last frame, or both. Use H3-Base-Ref2VA when the brief depends on several image, video, or audio references. A useful evaluation holds the creative goal stable while changing only the input mode, then records identity continuity, requested motion, camera logic, audio timing, text legibility, failure cases, and the amount of editorial repair required., and The repository section “Keep provider statements and observations separate” says: MiniMax lists advertising, branding, ecommerce, product design, interface design, and games among H3 scenarios. Those examples establish intended scope, not measured suitability for your campaign or pipeline. Save the exact prompt, authorized references, settings, output, reviewer, date, and acceptance rubric. Describe what happened in that run without turning a small sample into a universal quality, speed, cost, safety, or rights conclusion. Its FAQs narrow the reader's likely follow-up questions. For “What is MiniMax H3?”, the recorded answer is: MiniMax describes H3 as a general-purpose omni-modal generation system that understands text, images, video, and audio as context and generates video with native stereo audio. For “Does the open-weight release include the complete H3 system?”, the recorded answer is: No. MiniMax released H3-Base checkpoints, while its repository says H3-Context-IR and H3-Regenerate-2K are hosted modules and are not included in the initial open-weight release. For “Is MiniMax H3 available in SEELE?”, the recorded answer is: This page does not confirm that. Inspect the current SEELE Film & CG workspace and verify the visible model name, controls, region, entitlement, and terms before planning production. This is the complete page-specific editorial record available in the repository for “MiniMax H3.” It may define questions, cautions, or a review method, but it does not become external product evidence through repetition. Where one of these fields asks for a provider, model label, plan, capability, result, right, policy, or availability fact that no attached reference supplies, the only supported status is Unknown / not verified.

Model query guide: decision required by the taxonomy job

For “MiniMax H3,” the model orientation and fit review must identify whether the reader needs provider identity, a model family, an accepted input, a controllable behavior, an output constraint, an access surface, or a production-fit decision. The repository's intended workflow is to resolve identity, collect a bounded fact ledger, and use a representative authorized brief only for the workflow question documentation cannot settle. Capture exact label, provider, version, surface, region, date, inputs, controls, outputs, stated limits, failures, judgment, and unresolved questions. Apply that requirement only to the recorded topics “MiniMax H3 model”, “H3 video generation”, and “omni-modal video model”. Recognition, search demand, a showcase, or one successful result cannot establish current access, affiliation, quality, consistency, licensing, or SEELE support. Documented facts, direct observations, editorial judgments, and unknowns belong in separate fields: a fact retains a URL, publisher, claim scope, and date; an observation retains the exact product label, account surface, region, inputs, controls, failures, result, and observation date. A conclusion is bounded to the decision stated by this entry, not to every product that shares part of its wording. Refresh whenever the provider, version, inventory, interface, inputs, controls, output rules, plan, license, policy, or delivery requirement changes.

Model query guide: every literal qualifier in “MiniMax H3”

The wording triggers identity and workflow-fit wording. The phrase suggests a named model or product but supplies no verified provider, release, access, input, output, or production fact by itself. Resolve the provider-controlled identity and date, then document the accepted inputs, visible controls, output constraints, stated limits, and one bounded workflow test only when needed. Recognition of a name cannot establish identity, capability, availability, quality, price, license, or SEELE support. Read these qualifiers together with “MiniMax H3 model”, “H3 video generation”, and “omni-modal video model” and the model orientation and fit review; detaching one would answer a different query. A requested qualifier is not proof. Until a current claim-scoped source or authorized observation resolves it for the same provider, identity, version, surface, account, region, and date, record the result as Unknown / not verified and name the field that still needs checking.

Model query guide: profile facts or unresolved fields

The attached MiniMax H3 profile lists these inputs: Text, Optional first and last frame images, Image references, Video references, and Audio references. It records these documented capabilities: Text-to-audio-video, First-and-last-frame-to-audio-video, Multimodal reference-to-audio-video, Hosted Context-IR orchestration, and Separate 2K regeneration path. Its output notes are Official repository lists 4–15 second output, 24 FPS and 32 kHz stereo audio are documented, Several landscape, square, and portrait ratios are listed, and H3-Base produces the documented 768p path; 2K uses H3-Regenerate-2K. 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: Evaluate H3 when a shot depends on explicit relationships among visual, motion, performance, and audio sources. Record whether the test uses hosted Context-IR and regeneration, or only a local H3-Base checkpoint, because those paths are not equivalent. 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 “Multimodal relationship test”: Assign subject identity to an image, camera motion to a reference video, and vocal timing to authorized audio. Check whether each role remains distinct and whether the combined result preserves the intended priority., “First-to-last-frame test”: Provide a rights-cleared opening and ending composition with one motivated action and camera path. Review the transition for subject continuity, temporal logic, stereo sound, and unwanted scene invention., and “Text and audio-bearing test”: Brief a short scene with one legible text element, dialogue, ambience, and a timed effect. Inspect spelling, visual persistence, speech timing, channel balance, action synchronization, and the repair needed before delivery. Its recorded review checklist is “The exact H3 system path, interface, region, and date are recorded”, “Every reference has a stable label, authorized use, and one declared role”, “Identity, geometry, text, camera, timing, and audio are reviewed separately”, “Provider specifications are not rewritten as independent quality or commercial results”, and “SEELE availability, price, speed, and license fit are verified outside this page”. 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,” 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 model”, “H3 video generation”, and “omni-modal video model”, 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.

Model query guide: answer and refresh rule

An answer to “MiniMax H3” should state the requested decision, exact identity status, accepted and rejected evidence, evidence date, access context, any authorized observation, and every unresolved field. A proceed decision is limited to the documented provider, version, surface, account, region, inputs, controls, attempt allowance, and delivery target. A stop decision must identify its actual blocker: unresolved identity, absent source, unverified access, missing rights, unsupported input, failed output, policy risk, or poor workflow fit. Refresh whenever the provider, version, inventory, interface, inputs, controls, output rules, plan, license, policy, or delivery requirement changes. Also reopen the record when a cited page disappears, an identifier no longer matches the interface, a reference changes scope, or delivery creates a new rights, safety, disclosure, or technical condition. Until evidence supplies the missing fact, Unknown / not verified is more accurate than either a positive promise or a negative capability claim.