MiniMax H3 Chili
MiniMax's efficient multimodal video model — text, image, or reference to 4–15s clips with native stereo audio, from $0.12 per clip.
curl https://api.venice.ai/api/v1/video/queue \
-H "Authorization: Bearer $VENICE_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "minimax-h3-chili-text-to-video",
"prompt": "Aerial drone shot over a misty mountain valley at golden hour",
"duration": "5s",
"resolution": "768P",
"aspect_ratio": "16:9"
}'
# Use the returned queue_id with https://api.venice.ai/api/v1/video/retrieve.
# Call /video/complete after downloading if needed.Overview
What is MiniMax H3 Chili
MiniMax H3 Chili is MiniMax's multimodal video-generation model, offered on Venice in three modes: text-to-video, image-to-video, and reference-to-video. It renders 24 fps clips of 4–15 seconds at 768P with synchronized native stereo audio, and accepts up to 15 reference images, videos, or audio files for subject consistency.
Using it anonymously on Venice
On Venice, MiniMax H3 Chili runs under the anonymized privacy tier — your prompts and reference uploads are not stored, profiled, or used for training, and there is no personal generation history tied to your name. It's also uncensored: you describe the shot you want without a corporate filter second-guessing it. You pay per clip instead of a subscription, so a 4-second 768P draft costs $0.12 and scaling up is a pricing decision, not a plan upgrade.
Specifications
Datasheet
- Maker
- MiniMax
- Modality
- Text-to-video, image-to-video, reference-to-video
- Open weights
- No
- License
- Proprietary — no open weights
- Clip lengths
- 4s – 15s
- Resolutions
- 768P
- Mode
- text-to-video
- Audio
- Yes
- Prompt limit
- 7,000 chars
- Released
- July 31, 2026 (MiniMax H3 family); live on Venice since November 4, 2026
- Architecture
- Omni-modal generation stack — Contextual Omni Representation, H3-VAE, H3-Omni Transformer, In-Context Regeneration; 24 fps output
- Aspect ratios
- 16:9, 21:9, 4:3, 1:1, 3:4, 9:16
- Privacy on Venice
- Anonymized — prompts not stored
- Available on Venice since
- Nov 2026
Assessment
Strengths and limitations
- Three generation modes in one family: write a scene, animate a still image, or lock a character's identity with reference assets.
- Deep multi-reference control: references can carry separate roles for character identity, outfit, and environment instead of blending into one look.
- Native synchronized stereo audio: dialogue, ambience, and music generated with the picture, not bolted on afterward.
- Prompts are passed through verbatim with no silent rewriting, so what you write is what the model gets.
- Strong instruction following on complex multimodal direction, including precise cut timing stated in seconds.
- Cheapest entry point in MiniMax's H3 line on Venice — 768P clips start at $0.12 — and uncensored, with zero retention of your prompts.
- 768P only on Venice: the H3 family can reach higher resolutions elsewhere, but the Chili tier here caps at 768P.
- Runs a distilled 8-evaluation recipe, so it follows prompts less closely than full-step generation; vague prompts drift even more.
- No open weights: you cannot self-host, inspect, or fine-tune it.
- Third-party testing found facial consistency can weaken during motion at 768P, and higher-fidelity settings trade waiting time for sharpness.
- Free-text prompts sometimes come back with an unrequested music bed unless you use the structured prompt format.
Use cases
What it is good for
- 01Advertising and branding clips where a character, outfit, and setting must stay consistent across shots.
- 02E-commerce product video animated directly from still photography.
- 03Vertical social content — 9:16 and 3:4 aspect ratios with sound built in.
- 04Previsualization and animatics where cut timing is specified in the prompt.
- 05Character-driven shorts with dialogue or music generated alongside the visuals.
- 06Rapid 4-second drafts at $0.12 before committing to longer, more expensive renders.
Prompting
Getting better results
Describe the shot, not a request: subject, setting, time of day, weather, lighting, camera move, and motion. "Show me Christmas in Tokyo" drifts; "Shibuya crossing on Christmas Eve at night, light snow, crowds with umbrellas, slow cinematic dolly" lands.
Use the structured prompt format to control sound: state the soundscape, whether music plays, and whether anyone speaks — this cut unrequested music beds from 10 of 24 prompts to zero in one test bank.
In reference-to-video mode, the references fix who appears — spend your prompt words on setting, action, and camera move for them.
Be more detailed than feels necessary: the distilled 8-evaluation recipe follows vague prompts less faithfully than full-step generation.
State cut timings explicitly in seconds ("cut at 1.5s") — H3 is reported to honor stated cut points precisely.
Iterate at 768P · 4s for $0.12, then re-render the winning prompt at 15s once the framing is right.
Samples
Sample outputs
Generated on Venice with our standard prompt suite — the same prompts we run through every model of this type, so you can judge it like-for-like.
Slow aerial drone shot gliding over a misty mountain valley at golden hour, sunlight piercing clouds onto a winding river, ancient pine forests on either side, ultra-smooth motion, professional color grading, atmospheric haze, 4K
Calm ocean waves rolling onto a black-sand beach at sunrise, golden light on wet sand, foam dissolving into the shore, a single silhouetted figure at the waterline, smooth continuous forward push, serene cinematic atmosphere
A slow tracking shot through a rain-soaked Tokyo street at night, neon reflecting in puddles, steam rising from a food cart, a person with a translucent umbrella, shallow depth of field, teal-and-orange grade, smooth steady camera
Alternatives
How it compares
| Model | Best for | Max duration | Max resolution | Native audio |
|---|---|---|---|---|
| MiniMax H3 Chili | Multi-reference control on a budget | 15s | 768P | Yes |
| MiniMax H3 Max | Same family, higher fidelity | 15s | 1080P | Yes |
| Grok Imagine 1.5 | Fast stylized clips with sound | 15s | 1080p | Yes |
| Wan 2.7 Enhanced | Open-weights alternative | 15s | 1080p | Yes |
| Vidu Q3 | Longest single clip | 16s | 1080p | Yes |
MiniMax H3 Chili is the right pick when subject consistency across references and native audio matter more than raw resolution — and when per-clip cost does. Move to H3 Max or Grok Imagine 1.5 when 1080P is non-negotiable.
API
Call it from your code
Venice exposes this model through the REST API. Queue a generation with the model id.
curl https://api.venice.ai/api/v1/video/queue \
-H "Authorization: Bearer $VENICE_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "minimax-h3-chili-text-to-video",
"prompt": "Aerial drone shot over a misty mountain valley at golden hour",
"duration": "5s",
"resolution": "768P",
"aspect_ratio": "16:9"
}'
# Use the returned queue_id with https://api.venice.ai/api/v1/video/retrieve.
# Call /video/complete after downloading if needed.Pricing
What it costs on Venice
Pay per clip on Venice — price scales with resolution and duration (4s–15s), from $0.12.
New Venice accounts include a free daily allowance and 500 welcome credits — no credit card required.
FAQ
Frequently asked questions
MiniMax H3 Chili is a multimodal video-generation model from MiniMax, available on Venice in three variants: text-to-video, image-to-video, and reference-to-video. It generates 24 fps clips of 4–15 seconds at 768P with synchronized native stereo audio, and accepts up to 15 reference assets — images, video, or audio — for consistent subjects.
On Venice you pay per clip, with pricing that scales with resolution and duration. A 768P · 4-second clip starts at $0.12, and longer clips cost more. There is no subscription required — credits are consumed per generation.
MiniMax H3 Chili is not open source — the weights are closed and it runs as a hosted model on Venice, so it cannot be self-hosted or fine-tuned. It is free to try with Venice's welcome credits, and after that you pay per clip starting at $0.12.
Yes. The image-to-video variant (model id minimax-h3-chili-image-to-video) animates a still image into motion, with the same 4–15 second clip lengths, 768P resolution, and native stereo audio as the text-to-video variant.
Yes — that is the reference-to-video variant (minimax-h3-chili-reference-to-video). It accepts up to 15 assets, and images, videos, and audio can be combined: images guide subjects and style, videos guide motion, and audio guides voices and music. Audio cannot be used on its own; at least one image or video is required.
Yes. Every variant produces synchronized 32 kHz stereo sound natively — ambience, music, and speech. For the most control over the soundtrack, use the structured prompt format to explicitly state the soundscape, whether music plays, and whether anyone speaks; free-text prompts sometimes trigger an unrequested music bed.
Both are MiniMax H3-line video models with 15-second clips and native audio. H3 Max renders up to 1080P; H3 Chili caps at 768P on Venice but starts at $0.12 per clip. Use Chili for iteration and reference-driven work on a budget, and Max for final renders where resolution carries the shot.
It runs under Venice's anonymized privacy tier: prompts and reference uploads are not stored, profiled, or used for training, and generations are not tied to a personal history. It is also uncensored — you direct the scene without a content filter rewriting your intent.
On Venice, H3 Chili outputs 768P clips from 4 to 15 seconds at 24 fps, in six aspect ratios: 16:9, 21:9, 4:3, 1:1, 3:4, and 9:16. Prompts can run up to 7,000 characters.
Use MiniMax H3 Chili anonymously
Venice does not store your prompts. Chat history stays in your browser.