MiniMax H3 + FastH3 Turbo

About

MiniMax H3 creates complete short-form videos with picture and sound in one generation. Direct the action, camera, dialogue, ambience, effects, and music in the same prompt — no separate soundtrack pass required. It is a strong choice for dialogue scenes, brand films, product reveals, fashion clips, motion design, music visuals, and social content.

Start from a text prompt, animate an opening image, guide a transition with opening and closing frames, or use image, video, and audio references to shape the result. You can also pin up to eight keyframe images at exact moments in between, not only the first and last frame: the video lands on each one at its time, and a keyframe with a new angle, place, or light starts a new shot, so you can storyboard a whole clip. To keep a clip going, extend it: Text to Video picks the video up where it ends, carries its motion and sound into the new footage, and gives you the whole video or just the new part. Choose the workflow that matches how much creative control you want. Sound to Video also accepts an audio track on its own, with a starting image, or with both opening and closing frames.

Start with FastH3 Turbo for everyday creation. Built on FastVideo FastH3 4-step Preview v1 VSA DataFree, the FastVideo team's four-step distillation of MiniMax H3, it renders up to 2× faster than the LightX2V 4-step Turbo and up to 6× faster than Standard 768p H3 on a full 15-second clip. It is ideal for drafts, timing tests, quick iteration, and most social content, and its text and frame-guided workflows work with the same H3 LoRAs. Choose Standard H3 when you want the very best fine detail and audio polish. The familiar LightX2V Turbo stays one switch away, and multi-reference video is available on every tier, including FastH3 Reference.

Create videos from about 5 to 15 seconds in square, landscape, cinema, or portrait layouts, with dialogue, music, and sound generated together. Sogni offers 480p plus 544/768p-class output from the open-weights release, and FastH3 two-stage output (Beta) adds 2-stage 768p, 1080p, and 2K: a second pass refines the clip at twice its width and height with Sogni's in-house latent-enhance upscale, and 2K costs the same per second as Standard 768p. It is separate from MiniMax’s hosted 2K service. New: Reference to Video in 2K. Give MiniMax H3 up to 9 reference images, 3 reference videos and 3 audio clips and get a 2688 × 1536 clip back — the first pass on the Standard or Balanced Reference model, the enlargement refined by FastH3 — on ordinary 32 GB cards, with 2-stage 1080p (1920 × 1088) and 2-stage 768p beside it. 2-stage 768p is now the default for Standard and Balanced Reference to Video: it renders at half size and refines the enlargement with the same model, so a 15-second reference video renders on cards such as the RTX 5090 at the same price as 1-stage 768p. FastH3 Reference now runs Reference to Video on the FastH3 tier at every size, from one-stage 480p to 2-stage 2K. MiniMax reports stable dialogue support in 11 languages, with additional languages supported to varying degrees.

Sogni Web makes FastH3 the default Turbo engine for text-to-video, image-to-video, and first-and-last-frame video, with a switch back to LightX2V Turbo. The Sogni API accepts an exact model ID for every Standard, Balanced, LightX2V Turbo, FastH3 Turbo, FastH3 two-stage, FastH3 Reference, and two-stage Reference to Video workflow, and the Creative Agent Skill offers Standard, LightX2V Turbo, FastH3 Turbo, and FastH3 Reference selectors. Multi-reference video accepts up to nine images, three videos, and three audio clips on every tier, FastH3 Reference included.

Choose your H3 workflow and speed

Choose FastH3 Turbo for fast everyday creation — up to 6× faster than Standard — LightX2V Turbo for its familiar four-step look, or Standard when you want maximum detail and audio polish. Multi-reference video is available on every tier, including FastH3 Reference. Sound to Video uses FastH3 Turbo with your audio and optional opening and closing frames. Two-stage 768p, 1080p, and 2K output (Beta) is a Video Size choice in FastH3 settings and has its own six model IDs. Reference to Video goes up to 2K: two-stage output (Beta) is a Video Size choice on Standard, Balanced, and FastH3 — 2-stage 768p (the default), 1080p, and 2K (2688 × 1536) — with one more model ID per tier.

Optional MiniMax H3 video LoRAs

Open the H3 LoRAs panel to add realism, natural speech, better motion, fight choreography, faster or slower actions, a vintage look, tighter prompt adherence, or opt-in mature-theme knowledge without changing the underlying MiniMax H3 workflow.

• Realism People by fal — improves faces, skin texture, hands, lighting, and natural human motion.
• Natural Face & Speech by AdaptiveVision — makes people talking on camera look and sound more natural, with cheeks, brows, jaw, and lips moving together and clearer spoken English. Built for vlogs, podcasts, interviews, and presenters.
• Better Motion by AdaptiveVision — gives full-body movement more natural weight shifts, strides, turns, and gestures for dance, sport, and walking shots.
• Combat Base V2 by FourBunny — strengthens continuous fight exchanges, hit reactions, grappling, takedowns, collisions, and falls. Describe the fight as a step-by-step sequence of cause and effect.
• vh5tape Worn VHS by KennethFal — makes picture and sound look recorded off 1980s broadcast TV onto a worn VHS tape. Start the prompt with vh5tape.
• VBVR Video Reasoning by MisticRain69 — improves literal prompt adherence and ordered actions.
• Mystic X v5 by alcaitiff — adds opt-in mature-theme knowledge while preserving motion and fine detail. Version 4 remains available.
• Action Speed Slider by make676767 — positive strengths speed up punches, kicks, sword swings and other limb actions; negative strengths create slow motion.

Compatibility depends on the tier and workflow. Mystic X v5 supports multi-reference video, including FastH3 Reference; Mystic X v4 and Action Speed Slider support text, image and frame-guided video. Realism People uses two-stage FastH3 for text and image video. Worn VHS uses two-stage FastH3, LightX2V Turbo or Balanced for text and image video. VBVR Video Reasoning and Mystic X v5 appear after the Sensitive Content Filter is turned off. For a discreet overview of supported mature-theme workflows, see Sogni's uncensored AI video generator.

Prompting tips

We recommend leaving Sogni's AI Script Writer on. It is enabled by default in many Sogni interfaces and automagically turns a simple idea into a rich script tailored to H3. You can still add shot, camera, style, dialogue, and sound details whenever you want more control.

• Start with one sentence — With AI Script Writer on, a simple idea is enough. Say who or what is in the scene, what happens, and the vibe you want.
• Choose the right anchor — Use text alone for T2VA, an opening frame for I2VA, a closing frame for L2VA, both endpoints for FL2VA, or loose labelled image, video, and audio references for Ref2VA.
• Describe each keyframe — Keyframes steer the picture, so say what each one shows and when it lands. A keyframe with a new angle starts a new shot; a same-angle keyframe can arrive as a push-in instead of a cut.
• Preserve exact words — Put must-keep dialogue and visible text in your request exactly as written. If you ask for speech without writing a line, Script Writer can author one concise line; otherwise it should not invent speech.
• Direct sound as carefully as picture — Name ambience, physical effects, diegetic sound, and audience-only music separately because H3 generates native video and stereo audio together.
• Give every reference one job — For Ref2VA, say which asset controls identity, style, composition, motion, pacing, voice, ambience, or music, and state which source wins if they conflict.
• Let Script Writer place endpoint alignment — I2VA, L2VA, and FL2VA require MiniMax's exact opening line. It names the referenced picture, actual final shot, and snapped endpoint; T2VA and Ref2VA have no alignment line.
• Focus on one finished beat — H3 works best when each 5- to 15-second clip has a clear action, development, and payoff instead of trying to compress an entire story.

MiniMax H3 prompting guide

For everyday creation, leave Sogni's AI Script Writer on and write a natural brief. It expands the idea into MiniMax's model-native structure. Advanced callers can write that structure directly using the guide below; small formatting imperfections are worth correcting, but they should not replace the creative substance of a useful prompt.

1. Pick the input shape

• T2VA — text only; establish the whole scene.
• I2VA — one opening frame; continue forward from it.
• L2VA — one closing frame; infer a plausible earlier state and converge on it. Sogni routes this through the I2V model with a closing-frame role, not a separate L2V model ID.
• FL2VA — opening and closing frames; describe the continuous physical path between both endpoints.
• Ref2VA — loose labelled image, video, and audio references; assign every asset a specific job.

2. Use the three Base fields

integrated_multimodal_description: [Shot 1] ...
overall_soundscape: ...
non_diegetic_music: ...

[Shot 1] has no timestamp. Every later cut uses [Shot N] At MM:SS.mmm, ..., with contiguous shot numbers and strictly increasing times inside the rendered duration. Describe visible action, camera, dialogue, singing, diegetic music, and synchronized events in the main description. Put ambience, effects, and non-verbal human sound in overall_soundscape. Put only audience-only score in non_diegetic_music, or use N/A when there is no score.

3. Preserve speech and text precisely

Give each vocal source a stable ID such as (S1). Keep identity, action, and delivery outside the dialogue tag; put only the language and words inside <d>[English] Exact words.</d>. Preserve user-supplied dialogue character-for-character. If the request explicitly asks for speaking, dialogue, lyrics, or a vocal performance without supplying words, author one concise line; with no vocal intent, invent no speech. Use <scenetrans> at both connection points when one line crosses a cut and plain <cutoff> only when the video ending truncates speech. Never write tokenizer-internal <|...|> controls in a prompt, and never fix one by merely removing its pipe characters: caption markers become exact visible text in double quotes, lyrics markers become an ordinary <d>[Language] exact words</d> singing block, and <|cutoff|> becomes plain <cutoff>. Plain caption or lyrics boundary tags are not valid substitutes.

4. Add the endpoint alignment line

I2VA begins with the exact opening-frame alignment sentence. L2VA and FL2VA begin with their duration-aware alignment sentence naming the actual final shot and snapped endpoint. T2VA and Ref2VA use no alignment sentence. The Script Writer handles these lines automatically when it knows which frames are attached.

5. Use the six Ref2VA fields

subject_definitions:
summary:
retention_analysis:
detailed_description:
overall_soundscape:
non_diegetic_music:

Keep <Subject N>, <Picture N>, <Video N>, and <Audio N> meanings stable. The summary starts with task types chosen from keyframe completion, reference generation, video editing, video continuation, audio reuse, and audio reference; join multiple types with exactly + and never repeat one. Attached clips are loose references unless the runtime explicitly provides an edit or continuation relationship, so do not promise source-video transformation from file presence alone.

Use fully_preserved, partially_preserved, attribute_transfer, or weak_reference for visual retention; use fully_copy, partially_copy, reference, or weak_reference for audio. Bind a reference voice to its subject's (Sx) in subject_definitions, not in retention_analysis. A timbre, rhythm, emotion, or delivery reference does not authorize copying its words. Preserve explicitly reused speech and mark unintelligible spans [unclear] instead of guessing.

Ref2VA requires at least one image or video. It accepts up to nine images, three videos, three standalone audio clips, and 12 files total. Each reference video and audio clip must be 2–15 seconds; reference videos may total at most 15 seconds, and reference audio may separately total at most 15 seconds.

6. Stay within the open-weights model's real limits

H3 renders at fixed 24 fps on a 124 + n×17 frame grid, about 5.17–15.08 seconds, with a 7,000-character prompt limit. Preserve required fields, exact user text, reference jobs, retention markers, and shot timing before trimming redundant adjectives or audio prose. Sogni offers 480p plus 544/768p-class output from the open-weights release, FastH3 two-stage 768p, 1080p, and 2K output (Beta), and 2-stage 768p, 1080p, and 2K output for Ref2VA (Beta). Changing the prompt cannot enable an output resolution that is not yet open.

Source: MiniMax's official H3 prompt-writing skill, pinned to the reviewed open-weights revision.

Measured on one RTX 5090: the same 15-second 768p image-to-video prompt rendered in 2 min 6 s on FastH3 Turbo and 11 min 55 s on Standard, with Balanced and LightX2V Turbo in between. Every clip is embedded in FastH3 Speed Test: Four MiniMax H3 Tiers, One RTX 5090, One Cat in a Bath.

See Standard and Turbo side by side in the Sogni Engineering field guide, Your Prompt Is Now a Director.

See what it builds: Sogni Worlds makes playable point-and-click adventures where every crossing between pictures is a MiniMax H3 first-and-last-frame video. The agents that built it explain how in We Are the Agents Who Built Sogni Worlds.

See keyframes in action: Sogni Launches Multi-Keyframe MiniMax Generations pins stills at exact moments on Image, Sound and Reference to Video, and the lip sync holds through every cut.

Go fast with FastH3 Turbo

FastH3 Turbo is our go-to for fast everyday creation. It runs FastVideo FastH3 4-step Preview v1 VSA DataFree, the FastVideo team's four-step distillation of MiniMax H3, and renders up to 2× faster than the LightX2V 4-step Turbo and up to 6× faster than Standard 768p H3 on a full 15-second clip. Use it for drafts, timing tests, quick iterations, and most social content. Choose Standard H3 when maximum fine detail and audio polish matter most.

FastH3 is the default Turbo engine in Sogni Web for text, image, and first-and-last-frame video, with a switch back to the familiar LightX2V Turbo look. Its text and frame-guided workflows work with the H3 LoRAs. It costs 4 Spark ($0.02) per output second. Shorter clips see a smaller speed-up than 15-second clips. For references, use FastH3 Reference.

Faster Reference to Video with FastH3 Reference

FastH3 Reference brings Reference to Video to the FastH3 tier. It runs FastVideo's FastH3 Omni, a distilled MiniMax H3 reference-to-video model, with the same references and keyframes as every Reference tier. It renders one-stage 480p and 768p class, and 2-stage 768p, 1080p, and 2K (Beta); at 1080p and 2K, FastH3 Omni renders the first pass with all your references and FastH3 refines the enlargement. In Sogni Web, pick FastH3 in Reference to Video; it starts at 2-stage 768p.

Measured on one RTX 5090, model loaded: an 8-second 1344 × 768 clip with a keyframe rendered in 101 s on FastH3 Reference and 253 s on Balanced Reference. With three reference pictures, 2-stage 1080p took 351 s against 435 s, and 2-stage 2K 528 s against 577 s. FastH3 Reference costs 6 Spark ($0.03) per output second at 480p and 768p, 12 Spark ($0.06) at 2-stage 1080p, and 26 Spark ($0.13) at 2-stage 2K. Watch a 2-stage 1080p sample.

MiniMax H3 speed comparison

For 15-second 768p-class video, use Standard as the 1× baseline. Balanced runs the MiniMax H3 LightX2V 8-step recipe, Turbo runs the LightX2V 4-step recipe, and FastH3 Turbo runs FastVideo FastH3 4-step.

The times show relative render speed: if Standard takes 15 minutes, the faster tiers scale from that baseline. They exclude queue time and are not guarantees; workflow, input, worker hardware, and cache state can change results. For an illustrative 15.0-second cost comparison, 544/768p-class output is 240 Spark ($1.20) on Standard, 150 Spark ($0.75) on Balanced, 90 Spark ($0.45) on Turbo, or 60 Spark ($0.30) on FastH3 Turbo; actual billing follows the rendered frame duration, up to 15.08 seconds. At 480p, Standard costs 10 Spark ($0.05), Balanced 6 Spark ($0.03), and both Turbo engines 4 Spark ($0.02) per output second.

Measured, not illustrative. On a single RTX 5090 worker, the same 15.08-second 768 × 1024 image-to-video prompt rendered hot in 11 min 55 s on Standard, 5 min 30 s on Balanced, 3 min 0 s on LightX2V Turbo, and 2 min 6 s on FastH3 Turbo, with 3 min 39 s, 1 min 46 s, 1 min 7 s, and 55 s at 480p. Every clip is embedded so you can judge the quality trade yourself: read the FastH3 speed test.

Make a finished beat, not a silent motion test

MiniMax H3 creates video and sound together. Direct dialogue, ambience, effects, and music in the same brief, then get a complete clip without a separate soundtrack pass.

Choose the control your shot needs

Start from text, an opening frame, a closing frame, both endpoints, or a loose labelled reference set. Ref2VA on every tier, FastH3 included, can combine up to nine images, three videos, and three audio clips.

A powerful alternative for open creative work

MiniMax H3 is its own MiniMax model, not a Seedance derivative. Sogni offers 480p plus 544/768p-class output from the published weights, and FastH3 two-stage output (Beta) adds 2-stage 768p, 1080p, and 2K at the Standard 768p price. Sogni's in-house latent-enhance upscale is separate from MiniMax’s hosted 2K service. Creators 18 or older can opt into mature-content creation on Sogni, subject to Sogni's Terms of Use and the applicable MiniMax, LightX2V, and FastVideo licenses.

Where each workflow is available

Sogni Web offers T2VA, I2VA, L2VA, and FL2VA in Standard, Balanced, LightX2V Turbo, and FastH3 Turbo, plus Ref2VA in Standard, Balanced, LightX2V Turbo, and FastH3 Reference. Sound to Video supports audio alone, image + audio, and first-and-last-frame + audio through FastH3 Turbo. FastH3 two-stage 768p, 1080p, and 2K output (Beta) is a Video Size choice in FastH3 settings, including Sound to Video. Reference to Video delivers 2K: two-stage Reference to Video (Beta) is a Video Size choice on Standard, Balanced, and FastH3 — 2-stage 768p (default), 1080p, and 2K. The Sogni API accepts every workflow by exact model ID, and the Creative Agent Skill exposes Standard, Balanced, LightX2V Turbo, FastH3 Turbo, and FastH3 Reference selectors.

Simple base pricing

FastH3 costs 4 Spark ($0.02) per output second at every one-stage size, and 2-stage 2K costs 16 Spark ($0.08), the Standard 768p rate. Every tier, resolution and Reference to Video fee is in the pricing table, and Sogni shows the exact quote before rendering.

Sources and licenses

Review the official MiniMax H3 model card, MiniMax H3 Community License, official MiniMax H3 API pricing, LightX2V H3 Turbo model card, and LightX2V Turbo implementation, FastVideo FastH3 4-step Preview v1 VSA DataFree model card, FastVideo implementation, Hao AI Lab FastH3 preview post, Kijai INT8 ConvRot conversion used by Sogni, and the FastVideo FastH3 Omni model card for FastH3 Reference. Sogni has received written authorization from MiniMax to offer H3 through the platform.

FastH3 two-stage output: sizes and pricing

Two-stage output (Beta) renders a FastH3 clip, then refines it at twice the width and height with Sogni's in-house latent-enhance upscale, preserving the frame count, keyframes, and audio. Choose 2-stage 768p, 1080p, or 2K in the Video Size widget of FastH3 settings, including Sound to Video; the price class follows the delivered pixels.

Prices include FastH3 generation and the second-stage surcharge at the exact canvases shown: no surcharge for 768p, 6 Spark ($0.03) per second for 1080p, and 12 Spark ($0.06) for 2K. Other aspect ratios can fall into a different price class; Sogni shows the live quote before rendering. The one-stage “up to 6× faster” comparison above does not describe two-stage render times: a 2K clip takes roughly as long as Standard 768p.

The second stage is built on the open-source ComfyUI MiniMax H3 Latent Upscaler and its published weights. See the 2K samples, or check pricing for every generation option.

Two-stage Reference to Video (Beta)

Reference to Video has its own two-stage output on the Standard, Balanced, and FastH3 tiers, chosen in the same Video Size widget; 1-stage 768p stays available. On Standard and Balanced, 2-stage 768p renders at half size, then enlarges and refines the clip with the same model, so a 15-second reference video renders on 32 GB cards such as the RTX 5090 instead of waiting for the largest workers. 2-stage 1080p and 2-stage 2K render the first pass on the tier's model at the 768p canvas and hand the enlargement to FastH3 for refinement to 1920 × 1088 or 2688 × 1536; on Standard and Balanced they also render on 32 GB cards. LightX2V Turbo Reference is one-stage only.

2-stage 768p costs what 1-stage 768p does on the same tier. Reference video is 16 Spark ($0.08) per second at every 2-stage size, 2K included, since only the 768p first pass reads it. API model IDs: minimax-h3-ref2va-fp8_r2v_2stage (Standard), minimax-h3-ref2va-fp8_r2v_balanced_2stage (Balanced), and minimax-h3-fastvideo-omni-int8_r2v_turbo_2stage (FastH3); send the half-size render canvas from the table, such as 672 × 384 for 768p output.

Pricing

Priced per second of finished video, with sound included. FastH3 Turbo is the fastest and costs the least; 2-stage output renders smaller, then enlarges 2× and refines. Pay as you go (1 Spark = $0.005), or create with any tier on an Unlimited plan, subject to fair use.

2-stage 768p costs the same as 768p. *Reference to Video only.

Reference to Video inputs: each reference-video second costs 16 Spark ($0.08) at 768p and 10 Spark ($0.05) at 480p, the Standard output rate, on every tier; 2-stage 1080p and 2K output also bill 16 Spark ($0.08) per reference second. A 15-second reference plus 15 seconds of Standard 768p output bills like 30 seconds, 480 Spark ($2.40); Balanced and Turbo discount only the output seconds. Five reference images are included on Balanced and Standard, then 16 Spark ($0.08) each; audio is free.

The Relaxed network renders FastH3 Turbo at half price. Sogni shows the exact quote before rendering.

API

Start with a natural creative brief. Creative Agent expands it into the production prompt MiniMax H3 needs and runs the matching workflow.

const response = await fetch('https://api.sogni.ai/v1/chat/completions', {
  method: 'POST',
  headers: {
    'Content-Type': 'application/json',
    Authorization: `Bearer ${process.env.SOGNI_API_KEY}`,
  },
  body: JSON.stringify({
    messages: [{ role: 'user', content: "Create an 8-second MiniMax H3 video from this brief: A ceramic artist opens a glowing kiln as the camera slowly pushes in; fire crackles, tools clink softly, and a restrained string score begins" }],
    sogni_tools: 'creative-agent',
    sogni_tool_execution: true,
  }),
});

if (!response.ok) throw new Error(await response.text());
const { choices } = await response.json();
console.log(choices[0].message.content);

curl https://api.sogni.ai/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $SOGNI_API_KEY" \
  -d '{
    "messages": [{ "role": "user", "content": "Create an 8-second MiniMax H3 video from this brief: A ceramic artist opens a glowing kiln as the camera slowly pushes in; fire crackles, tools clink softly, and a restrained string score begins" }],
    "sogni_tools": "creative-agent",
    "sogni_tool_execution": true
  }'

Creative Agent accepts the short brief above and prepares MiniMax H3's structured production prompt. Calling a worker directly? Follow the MiniMax H3 integration guide and use one of these exact model IDs: minimax-h3-fl2va-fp8_t2v, minimax-h3-fl2va-fp8_i2v, minimax-h3-fl2va-fp8_flf2v, minimax-h3-ref2va-fp8_r2v, minimax-h3-ref2va-fp8_r2v_2stage, minimax-h3-ref2va-fp8_r2v_balanced_2stage, minimax-h3-fastvideo-int8_t2v_turbo, minimax-h3-fastvideo-int8_i2v_turbo, minimax-h3-fastvideo-int8_flf2v_turbo, minimax-h3-fl2va-fp8_t2v_turbo, minimax-h3-fl2va-fp8_i2v_turbo, minimax-h3-fl2va-fp8_flf2v_turbo, minimax-h3-ref2va-fp8_r2v_turbo, minimax-h3-fastvideo-omni-int8_r2v_turbo, minimax-h3-fastvideo-omni-int8_r2v_turbo_2stage, minimax-h3-fastvideo-int8_a2v_turbo, minimax-h3-fastvideo-int8_ia2v_turbo, minimax-h3-fastvideo-int8_flfa2v_turbo, minimax-h3-fastvideo-int8_t2v_turbo_2stage, minimax-h3-fastvideo-int8_i2v_turbo_2stage, minimax-h3-fastvideo-int8_flf2v_turbo_2stage, minimax-h3-fastvideo-int8_a2v_turbo_2stage, minimax-h3-fastvideo-int8_ia2v_turbo_2stage, minimax-h3-fastvideo-int8_flfa2v_turbo_2stage. Full reference at docs.sogni.ai.