Primary sources last checked on 13 Nov 2026. Verification freshness is tracked separately from factual change, and unknown values remain explicit rather than inferred.
NotableCurrent first-place open-weight text-to-video model on the Artificial Analysis video leaderboard at review time.Curation method →
Open Model Weights records MiniMax-H3 as a 33.123B-parameter image text to video open-weight model in the MiniMax-H3 family, with MiniMax listed as the developer. The checked source repository provides 112 recognized weight artifacts, format evidence for Safetensors and a context window that is not declared in the checked standard sources. Values that are not evidenced in the checked sources remain explicitly undisclosed or unknown.
The repository declares MiniMax H3 Community License Agreement. Territory-restricted: the standard grant EXCLUDES the EU, UK, South Korea and US; separate authorization is required there. Commercial use in permitted territories carries license, safeguards, display and revenue-threshold conditions.. Open Model Weights reports the checked license evidence and classification; this is not legal advice.
No base model is declared in the checked standard lineage metadata, so Open Model Weights does not infer one. Training recipe status: training-described-in-model-card; data disclosure status: not-disclosed-in-standard-metadata-or-obvious-model-card-section.
MiniMax-H3 is multimodal audio-video generation family with image/text/video input signals. OMW checked the public source repository for 33.1B observed published parameter values, available weight artifacts without a verified text-style context window, and a other license declaration. More than a hundred recognized weight artifacts are present in the checked repository, and the metadata-derived parameter count differs in scope from isolated model-card parameter mentions. OMW does not convert either number to observed frame rate or audiovisual synchronization accuracy. Its Notable badge is a separate editorial discovery signal, not a performance measurement by OMW.
Primary technical focusmultimodal audio-video generation family with image/text/video input signalsSource ↗
Weight metadata33.1B observed published parameter valuesSource ↗
Checked context evidenceNot disclosed as a text-model context length in checked sourcesSource ↗
Runtime evidence classPublisher/source mentions only; no OMW execution testSource ↗
WHY IT MATTERS
The repository exposes tasks spanning text-to-video, image-to-video and synchronized audio-video generation. These source labels describe intended workflows, not a guarantee that one installed interface implements every combination.
For video generation, testing should cover temporal continuity, audio alignment, output duration and motion fidelity. The model should not be ranked by text LLM benchmarks simply because it has billion-class parameters.
MODEL-SELECTION VIEW
Strong fit when…
For video generation, testing should cover temporal continuity, audio alignment, output duration and motion fidelity. The model should not be ranked by text LLM benchmarks simply because it has billion-class parameters.
Evaluations that require multimodal audio-video generation family with image/text/video input signals and an inspectable original weight/source revision.
Important considerations
The official license is custom/other. Verify commercial rights and content restrictions directly, especially when generated audio and video are used for customers.
Compare fixed prompts and reference inputs at identical duration and resolution, report inference time and audio-video sync separately, and retain the exact model revisions.
Architecture and task-specific behavior
The repository exposes tasks spanning text-to-video, image-to-video and synchronized audio-video generation. These source labels describe intended workflows, not a guarantee that one installed interface implements every combination.
More than a hundred recognized weight artifacts are present in the checked repository, and the metadata-derived parameter count differs in scope from isolated model-card parameter mentions. OMW does not convert either number to observed frame rate or audiovisual synchronization accuracy.
Verified configuration versus usage claims
The primary registry record identifies MiniMax-H3 as Image Text To Video. No text-style maximum context is established in the checked source record. Do not import context lengths from unrelated language models into this modality. This page treats a checked config field, the model-card prose and a deployment benchmark as three distinct kinds of evidence.
In the current checked snapshot, 33.1B observed published parameter values are recorded with a source repository revision. Those weight-parameter values do not determine how much GPU memory, active compute or KV-cache is required by a real application; sparse activation, quantization and runtime overhead each need separate verification.
What the actual weight repository contains
OMW checked 112 recognized weight artifacts in Safetensors format. The 33.1B observed published parameter values come from the published repository data, not from a newly executed OMW performance test. A file count is not a checksum audit, a full weight-binary comparison or a statement about throughput.
The source revision and exact listed weight files can be inspected through the linked OMW evidence API and the official model repository. Before distributing or deploying a model, pin the precise commit and independently validate downloaded artifact checksums where available. An upstream revision change must not automatically be described as changed weight bytes.
License and intended commercial deployment
The model's license is recorded as other, with commercial use unknown or requiring individual assessment. Public downloadable files do not confer permission for every commercial deployment.
The official license is custom/other. Verify commercial rights and content restrictions directly, especially when generated audio and video are used for customers. Keep publisher license declarations separate from third-party interpretations, custom usage agreements and any deployment-specific regulatory obligations. Where a permission is unknown, the registry intentionally reports unknown rather than guessing a permissive answer.
Runtime, hardware and reproduction limits
The source record includes minimax-h3, ComfyUI, Diffusers, SGLang, vLLM as declared libraries or model-card mentions. This is provenance-aware compatibility information only: OMW has not executed a model load, generated outputs or benchmarked the exact checkpoint on a specified GPU. Runtime support for a particular version, precision and processor remains to be tested.
Compare fixed prompts and reference inputs at identical duration and resolution, report inference time and audio-video sync separately, and retain the exact model revisions. The generic weight-only memory estimate in the model snapshot does not include multimodal processors, activation buffers, temporary tensors, cache growth, device placement or kernel availability; these can dominate real deployments.
When to choose this model and how to compare it
For video generation, testing should cover temporal continuity, audio alignment, output duration and motion fidelity. The model should not be ranked by text LLM benchmarks simply because it has billion-class parameters.
For a fair comparison, choose another model from the same task family and freeze the task dataset, prompt or preprocessing settings, tested checkpoint revisions and compute budget. The official license is custom/other. Verify commercial rights and content restrictions directly, especially when generated audio and video are used for customers. A Leading or Notable label remains a discovery cue rather than a universal model ranking or an OMW benchmark result.
IMPLEMENTATION & EVALUATION
How to evaluate MiniMax-H3 before deployment
Use these as proposed acceptance tests rather than OMW-confirmed performance results. The objective is to establish whether the exact MiniMax-H3 revision is legal to use and deployable for a specific workload.
STEP 01
Freeze model artifacts and license evidence
Record the official MiniMax-H3 repository SHA, recognized Safetensors artifacts and the other license evidence before attempting an installation.
REPRODUCIBLE ACCEPTANCE TEST
An immutable source revision, artifact manifest, checksums where available and dated license record are saved together.
WHAT CAN GO WRONG
Mutable model branches or unreviewed third-party weight conversions can silently change the experiment or applicable terms.
STEP 02
Reproduce the documented task interface
The repository exposes tasks spanning text-to-video, image-to-video and synchronized audio-video generation. These source labels describe intended workflows, not a guarantee that one installed interface implements every combination.
REPRODUCIBLE ACCEPTANCE TEST
The intended Image Text To Video workload runs under a documented processor or tokenizer path with exact dependency versions, or the failure is recorded.
WHAT CAN GO WRONG
The official license is custom/other. Verify commercial rights and content restrictions directly, especially when generated audio and video are used for customers.
STEP 03
Measure deployment and model-specific outputs
Compare fixed prompts and reference inputs at identical duration and resolution, report inference time and audio-video sync separately, and retain the exact model revisions.
REPRODUCIBLE ACCEPTANCE TEST
Raw evaluation inputs, response or media outputs, memory high-water mark, request latency and error logs are retained and the environment can be reconstructed.
WHAT CAN GO WRONG
Publisher benchmark claims and parameter-derived estimates are not interchangeable with measurements on the chosen hardware.
STEP 04
Compare against a relevant verified alternative
For video generation, testing should cover temporal continuity, audio alignment, output duration and motion fidelity. The model should not be ranked by text LLM benchmarks simply because it has billion-class parameters.
REPRODUCIBLE ACCEPTANCE TEST
Run the same tasks and quality rubric against LTX-2.5 and Qwen-Image-2.1, with source revisions and test limitations disclosed.
WHAT CAN GO WRONG
A different license, modality, processor or sampling budget makes a naive head-to-head score misleading.
MiniMax-H3 is multimodal audio-video generation family with image/text/video input signals. The repository exposes tasks spanning text-to-video, image-to-video and synchronized audio-video generation. These source labels describe intended workflows, not a guarantee that one installed interface implements every combination.
What is the verified context length for MiniMax-H3?
No text-style maximum context is established in the checked source record. Do not import context lengths from unrelated language models into this modality.
How big are the published MiniMax-H3 weights?
33.1B observed published parameter values; 112 recognized repository weight entries (Safetensors). This is a weight-metadata observation, not a GPU memory measurement.
Can MiniMax-H3 be used commercially?
The model's license is recorded as other, with commercial use unknown or requiring individual assessment. Public downloadable files do not confer permission for every commercial deployment.
Which inference runtimes support MiniMax-H3?
The checked model documentation mentions minimax-h3, ComfyUI, Diffusers, SGLang, vLLM. OMW has not run this checkpoint under those runtimes; compatibility must be reproduced for the selected revision.
How should MiniMax-H3 be benchmarked?
Compare fixed prompts and reference inputs at identical duration and resolution, report inference time and audio-video sync separately, and retain the exact model revisions. Keep environment, source commit, data and evaluation method fixed, and distinguish publisher results from your own repeatable tests.
03 · PRACTICAL TECHNICAL PROFILE
Deployment, licensing and reproducible evidence.
Inspect weight-memory estimates, model lineage and runtime evidence separately from promotional or untested performance claims.
DEPLOYMENT PROFILE
Weight memory, without pretending it is a deployment guarantee.
These bars visualize weight-only decimal-GB estimates. They exclude KV cache, activations, runtime overhead, optimizer state and sharding.
WEIGHT-ONLY MEMORY
BF16 / FP16~66.2 GB
FP8 / INT8~33.1 GB
INT4~16.6 GB
FIELD EVIDENCE MAP
WeightsVerified
LicenseVerified
Commercial useUnknown
ContextNot declared
FormatsVerified
LineageNot declared
Training recipeTraining described in model card
Training dataNot disclosed in standard metadata or obvious model card section
RuntimeSource-derived · not runtime-tested
LICENSE & USE
MiniMax H3 Community License Agreement
Territory-restricted: the standard grant EXCLUDES the EU, UK, South Korea and US; separate authorization is required there. Commercial use in permitted territories carries license, safeguards, display and revenue-threshold conditions.
Declared licenseMiniMax H3 Community License Agreement
ClassificationUnknown
VerificationLicense file checked
This section summarizes the checked repository evidence. It is not legal advice. Inspect license evidence ↗
LINEAGE & PROVENANCE
What the checked sources disclose — and what they do not.
Open Model Weights does not infer a hidden base model or training corpus when the checked sources do not declare one.
LINEAGE
BASE / PARENTNot declared in model-card metadataUnknown
not inferred
CURRENT RECORDMiniMax-H3MiniMax-H3
TRAINING DISCLOSURE
Training recipeTraining described in model card
Training dataNot disclosed in standard metadata or obvious model card section
FamilyMiniMax-H3
RUNTIME SIGNALS
Compatibility claims stay attached to their evidence level.
Declared and model-card-mentioned runtimes are discovery signals. They are not treated as Open Model Weights runtime tests unless explicitly measured.
Declaredminimax-h3
MentionedComfyUI
MentionedDiffusers
MentionedSGLang
MentionedvLLM
OMW runtime testNot performed
EVIDENCE TRAIL
Sources and observed history.
This is where the article remains auditable: source links on one side, the registry's observed change history on the other.
Three evidence-linked visual summaries: parameter scale, illustrative precision memory and verifiable provenance. No performance ranking is implied.
01 / PARAMETER SCALE
Total versus active parameters
Total parameters33.12B
Active parameter count not established.
Repository-derived or publisher-described parameter count
Published total parameters are not a model-quality or performance score.
02 / MEMORY ESTIMATE
Weight-only precision footprint
BF16 / FP1666.25 GB
FP8 / INT833.12 GB
INT416.56 GB
Hypothetical weight-only decimal GB, not measured peak VRAM or guaranteed support for each format. KV cache, activations, runtime and sharding are excluded.
Repository SHA 0919fb684102…; a source check is not an inference benchmark or legal clearance. See the per-model evidence JSON for the exact source record.
A verification check does not pretend the model changed. Open Model Weights updates the verification date after a real primary-source check, while factual-update dates move only when retained source evidence changes.
No active conflict · 1 comparison signalSource values differ in scope, granularity or claim strength; OMW does not overstate those differences as contradictions.
Comparison signals are not conflicts.1 current signal keeps scope/granularity differences visible without asserting contradiction.
COMPARISON SIGNALtotal parameters billionsscope or granularity difference
Declared licenseMiniMax H3 Community License Agreement
License IDminimax-h3-community-license
Commercial useTerritory-restricted: the standard grant EXCLUDES the EU, UK, South Korea and US; separate authorization is required there. Commercial use in permitted territories carries license, safeguards, display and revenue-threshold conditions.
Repository license filesLICENSE
Classification describes the checked license terms; it is not legal advice.
Weight-only decimal-GB estimate from total parameter count; excludes KV cache, activations, runtime overhead, optimizer state and sharding.
04 · FORMATS
Formats & precision evidence
Safetensors
BF16: not observedFP16: not observedFP8: not observedINT8: not observedINT4: not observedGGUF: not observed
“Not observed” means no matching repository artifact/name/dtype signal was found; it does not prove no third-party conversion exists.
05 · LINEAGE
Base model & family
FamilyMiniMax-H3
VariantMiniMax-H3
Declared base modelNot declared in model-card metadata
RelationNot declared
06 · TRAINING ASSETS
Recipe & data disclosure
Training recipetraining-described-in-model-card
Data disclosurenot-disclosed-in-standard-metadata-or-obvious-model-card-section
EvaluationsNot classified by automated verifier
Intermediate checkpointsNot classified by automated verifier
07 · RUNTIME SUPPORT
Declared, mentioned and externally evidenced runtimes
Declared libraryminimax-h3
Declared/taggedminimax-h3
Mentioned in model cardComfyUI, Diffusers, SGLang, vLLM
Artifact signalsNone
External primary runtime evidenceNone
Tested by Open Model WeightsNo — source-derived support only
Open Model Weights separates model-repository evidence from external runtime-release evidence. An external support statement does not change the official model repository's weight formats and does not imply an Open Model Weights inference benchmark.
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Repository revision check
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Repository revision check
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Field-by-field verification
Checked Hugging Face API metadata, exact repository file list, model card and config when accessible.
Repository revision check
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Repository revision check
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Repository revision check
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Repository revision check
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Repository revision check
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Repository revision check
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Repository revision check
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Repository revision check
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Repository revision check
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Repository revision check
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Field-by-field verification
Checked Hugging Face API metadata, exact repository file list, model card and config when accessible.
Repository revision check
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Repository revision check
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Repository revision check
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Repository revision check
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Repository revision check
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Repository revision check
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Repository revision check
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Repository revision check
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Repository revision check
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Repository revision check
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Repository revision check
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Repository revision check
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Repository revision check
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Repository revision check
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Repository revision check
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Repository revision check
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Repository revision check
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Repository revision check
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Repository revision check
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Repository revision check
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Field-by-field verification
Checked Hugging Face API metadata, exact repository file list, model card and config when accessible.
Repository revision check
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Repository revision check
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Repository revision check
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Repository revision check
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Repository revision check
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Field-by-field verification
Checked Hugging Face API metadata, exact repository file list, model card and config when accessible.
Registry record created
Previous verification level: discovered-pending.
licenseverified
commercial useunknown
contextnot-disclosed
weight filesverified
formatsverified
base modelnot-declared
training recipetraining-described-in-model-card
data disclosurenot-disclosed-in-standard-metadata-or-obvious-model-card-section
runtime supportsource-derived-not-runtime-tested
MODELS FOR AI AGENTS · ROLE EVIDENCE
Where this model fits in an agent system
Video generation. A potential media-processing tool invoked by a separate controller; image/video/speech generation is not agent autonomy.
The curated notable tier is not an agent-readiness rating. Model-specific tool-call success, long-horizon task accuracy and compatible runtime configurations are not independently established here.
Source-linked license: MiniMax H3 Community License Agreement (unknown commercial-use classification; see original terms).
Context: Unknown or not declared.
BF16 weights: 66.25 estimated decimal GB, not runtime VRAM.