Kimi K3 + MiniMax H3 Max: Video Prompt Workflow, Costs & Examples

Kimi K3 cannot render video, but it makes a very good video director. The workflow that works in September 2026: K3 turns a rough idea into a timed, shot-by-shot prompt in the structure MiniMax’s own guide asks for, MiniMax H3 Max renders it as a 5 to 15 second clip at $0.05 to $0.08 per second, and K3 watches the result through its video input before the next take. A 10-second 768P clip costs $0.80; the K3 side adds a few cents.

This guide reflects both models as of September 25, 2026. The four example prompts are original illustrations of the format, not claimed results.

What MiniMax H3 Max actually is

The naming is messy, so here is the family tree from primary sources:

• MiniMax H3 (also sold as Hailuo 3) was released July 31, 2026. It takes text, images, video and audio in and returns video with native stereo sound at up to 2K and 15 seconds. The model card lists a 33B-parameter dense transformer, 24 FPS, 32 kHz stereo audio, six aspect ratios from 21:9 to 9:16 and stable dialogue in 11 languages. The weights use the MiniMax H3 Community License, which requires an application for users in the USA, EU, UK and South Korea.
• H3 Max is not a MiniMax in-house tier. It is a post-trained version of H3 built by fal, announced August 27, 2026, with added training data aimed at prompt adherence and visual quality. fal claims a 5-second clip in about 3 seconds, “roughly 35x the throughput of the official MiniMax H3 endpoint.”
• Where to call it: MiniMax’s API as MiniMax-H3-Max, fal, Vercel AI Gateway (minimax/minimax-h3-max, added August 30), OpenRouter (minimax/hailuo-3-max) and Runware.

fal’s September 1 press release cites #1 on Design Arena’s image-to-video leaderboard (Elo 1,341) and #1 on Artificial Analysis (Elo 1,201 across 2,177 samples). The leaderboards are real, but the figures come from the vendor’s announcement and arena ranks move weekly.

H3 vs H3 Max at a glance

	MiniMax H3	H3 Max (by fal)
Released	July 31, 2026	August 27, 2026
Duration	4–15 s (MiniMax API)	5–15 s
Resolution	768P / 2K	480P / 768P (fal also offers 1080P)
Frame rate	24 FPS	24 FPS (fal default 1344x768)
Audio	Native stereo, same pass	Synchronized audio, same pass
Inputs	Text, first/last frame, multi-reference (up to 9 images, 3 videos, 3 audio)	Text, first frame, optional last frame; references vary by provider
MiniMax price, 768P	$0.08/s	$0.08/s
MiniMax price, other tier	2K $0.13/s	480P $0.05/s
Prompt limit	7,000 characters	7,000 characters

Sources: MiniMax pricing and video generation docs, the H3 model card, fal’s H3 Max page.

One conflict: MiniMax’s docs say both models support reference generation, while Vercel describes H3 Max as taking “a text prompt or a starting image.” If you need multi-reference input with Max, test it on your provider first.

Why use Kimi K3 as the writer

Any chat model can write a video prompt. K3 has four properties that suit this job:

1. It holds a whole production in context. The 1,048,576-token window fits a character bible, style guide, a 40-shot storyboard and every prompt already rendered. Character drift between shots is the most common multi-clip failure, and keeping locked descriptions in one context is the cheapest fix. (See the K3 specifications.)
2. Caching makes that context cheap. Cached input is $0.30 per million tokens versus $3 uncached, so a 30,000-token project bible costs under a cent per call after the first.
3. It sees images and video. K3 can study a product photo before writing an image-to-video prompt, then watch the clip H3 Max produced. More in the K3 vision guide.
4. It follows structure. H3’s format means named sections, timestamped cuts, speaker tags and dialogue markup. That is formatting work, and K3 does schema-shaped output reliably.

The prompt structure H3 responds to

MiniMax published a video prompt writing guide alongside the H3 weights. Since H3 Max is post-trained from H3, it is the right starting format for Max too:

Element	What the guide asks for
integrated_multimodal_description	Visuals, action, dialogue and on-screen sounds, in chronological order
overall_soundscape	1–4 sentences: ambience, footsteps, impacts, breathing. No dialogue or music
non_diegetic_music	1–3 sentences on instruments, tempo, rhythm, dynamics, not mood words. “N/A” if none
Camera motion	Type (push, pan, truck, tilt, pedestal, arc, tracking, static, shake, POV, roll, zoom) plus amplitude plus speed, woven into the prose
Shot cuts	No timestamp on shot 1; then [Shot 2] At 00:03.500, the camera cuts to...
Speakers	Stable IDs (S1), (S2); voice described outside the tags; lines in <d>[English] ...</d>
On-screen text	In quotation marks, original language, never translated
Image-to-video	An alignment line first: Picture 1 is fully referenced at 0.00 seconds

Two notes. fal’s own H3 Max advice is looser: stay at 768P, generate 5 to 10 seconds, and “describe the sound as well as the shot, since audio is generated in the same pass as the picture.” And the three field names are the guide’s structure, not a documented JSON API schema. I have K3 emit JSON so a script can validate it, then flatten it into three labeled text sections for the renderer.

Step-by-step: the K3 to H3 Max workflow

1. Write a one-paragraph brief. Purpose, audience, length, aspect ratio, and the one image people should remember.
2. Ask K3 for a shot list first. Three to six shots with durations. Approve or edit it before any prompt writing; this is where the creative decisions live.
3. Lock a bible. Have K3 write exact clothing, colors, materials and lighting for every recurring subject, and reuse it in every request.
4. Generate each render prompt with the system prompt below. Keep reasoning_effort at low or high; max effort is slow at around 34 output tokens per second and rarely improves a 400-word prompt.
5. Draft at 480P, 5 seconds ($0.25 a take). Fix the seed while testing wording, so you compare prompts rather than luck.
6. Review with K3’s video input (below) and change one variable at a time.
7. Promote winners to 768P at full length; use regular H3 at 2K only where needed.
8. Assemble in a real editor. Neither model cuts your final piece.

A director system prompt for K3

You are a film director writing prompts for the MiniMax H3 / H3 Max
video model. Output valid JSON only, with keys: "task_prefix",
"integrated_multimodal_description", "overall_soundscape",
"non_diegetic_music", "duration_s", "aspect_ratio", "char_count".

Rules:
- Total prompt text under 3,500 characters (renderer limit is 7,000).
- Chronological order. No timestamp on the first shot. Later shots:
  "[Shot N] At 00:SS.mmm, the camera cuts to ...".
- Every camera move states type, amplitude and speed.
- Speakers get stable IDs (S1), (S2); describe the voice outside the tag;
  lines go inside <d>[English] ...</d>; never alter user-given lines.
- Quote on-screen text exactly.
- overall_soundscape: 1-4 sentences, no dialogue, no music.
- non_diegetic_music: 1-3 sentences on instruments, tempo, dynamics, or "N/A".
- Reuse CHARACTER BIBLE wording verbatim for recurring subjects.
- No logos, real people or copyrighted characters.
- Image-to-video: task_prefix states Picture 1 is fully referenced at
  0.00 seconds. Otherwise task_prefix is "".

The 3,500-character cap is my choice, not a MiniMax rule. K3 runs verbose, and longer prompts give the renderer more details to drop.

Four example prompts (original, for illustration)

These show the format as flattened text you would paste into a renderer. No render quality is claimed.

Example 1: Text-to-video, one shot, 5 s, 16:9

integrated_multimodal_description: A bicycle courier in a yellow rain jacket and black helmet rides toward the camera down a narrow wet street at night, neon signs reflecting in puddles. The camera tracks backward with small amplitude at medium speed, keeping the rider centered at waist height. The front wheel cuts through a puddle and throws a low spray. The rider glances up at a sign reading “OPEN 24H” and pedals harder.

overall_soundscape: Steady rain on pavement and metal awnings. A bicycle chain ticking and tires hissing through water. One distant car horn.

non_diegetic_music: N/A

One subject, one camera move, one beat of action. Five seconds holds about one idea.

Example 2: Image-to-video from a product photo, 6 s, 1:1

task_prefix: For the target video, at 0.00 seconds into the target video, Picture 1 (from Shot 1) is fully referenced.

integrated_multimodal_description: A steel field watch with a green dial rests on a slab of dark slate, lit by hard light from the upper left. The camera arcs right with small amplitude at slow speed, and a thin line of light travels across the crystal. The second hand sweeps smoothly. A single drop of water rolls off the edge of the slate and falls out of frame.

overall_soundscape: Quiet room tone. The soft mechanical tick of the watch, close to the microphone. One faint drip.

non_diegetic_music: A single low piano note held, then a second note a fifth higher, slow tempo, very soft.

Have K3 look at the actual photo first and describe only what is in it. Details the image contradicts are a common cause of warping.

Example 3: Two-shot dialogue with a cut, 10 s, 16:9

integrated_multimodal_description: In a narrow secondhand bookshop lit by warm lamps, (S1), a woman in her sixties with short grey hair and a navy cardigan, stands behind the counter. (S2), a young man in a brown corduroy jacket, sets a thick paperback down. The camera is static at medium distance. (S2), in a hesitant, quiet voice, says [English] Do you have the second volume? [Shot 2] At 00:05.000, the camera cuts to a close-up of (S1). She smiles slightly and, in a dry, low voice, says [English] Nobody has ever asked me that. The camera pushes in with small amplitude at slow speed.

overall_soundscape: Muffled street noise through a closed door. The dull thud of the book on wood. Pages rustling nearby.

non_diegetic_music: N/A

Two short lines in 10 seconds leaves room for reactions. Four lines is the usual mistake.

Example 4: Three-shot product teaser, 15 s, 9:16

integrated_multimodal_description: Close-up of a matte white ceramic kettle on a gas stove, a small blue flame beneath it, morning light through a window behind. The camera is static; steam begins to rise from the spout. [Shot 2] At 00:05.000, the camera cuts to a top-down shot of hot water pouring into a glass teapot of loose green tea leaves that swirl and unfurl; the camera pedestals down with small amplitude at slow speed. [Shot 3] At 00:10.000, the camera cuts to a medium shot of a hand setting the teapot on a wooden table beside an open notebook; the camera pushes in with medium amplitude at slow speed until the printed text “Slow mornings” fills the upper third of the frame.

overall_soundscape: A low gas hiss rising into a soft kettle whistle. Water pouring and gentle clinks of glass. A cup set down on wood.

non_diegetic_music: Fingerpicked acoustic guitar at a slow, even tempo, entering softly after the first shot and staying under the effects.

Clean 5-second cut points, one camera move per shot, and short quoted text, because long rendered text is still where video models fail most.

Cost math: per clip and per project

H3 Max bills per output second; Kimi K3 bills per token. Rates are MiniMax’s pay-as-you-go card and Moonshot’s official K3 pricing.

Clip	H3 Max 480P ($0.05/s)	H3 Max 768P ($0.08/s)	H3 2K ($0.13/s)
5 seconds	$0.25	$0.40	$0.65
10 seconds	$0.50	$0.80	$1.30
15 seconds	$0.75	$1.20	$1.95

Input extras on MiniMax: for H3 Max, audio references are free, the first 2 images are free and extra images cost $0.074 each (H3: first 5 free, then $0.04). fal lists H3 Max at the same $0.05/$0.08 plus $0.16/s for 1080P, and its model page showed a 50% launch discount through September 30 when I checked. Vercel ran its own 50% promotion from August 30 to September 13. Check the live price.

The K3 side, estimated. A typical render-prompt call is about 3,000 tokens of system prompt plus bible (cached after the first call), 300 new tokens of brief, and roughly 4,000 output tokens including reasoning. That is about $0.06 of output and under $0.01 of input, cheaper than one second of 480P video. These token counts are my estimates; check your dashboard.

Review calls. Moonshot counts image and video tokens dynamically from resolution and keyframes and offers an estimate-tokens API, but publishes no flat per-clip figure. As a scale check only: even 20,000 input tokens for a clip is $0.06 uncached.

A worked budget: one 30-second vertical ad

Stage	Quantity	Cost
K3: shot list, bible, 3 render prompts, revisions	~15 calls	~$1.00 (est.)
H3 Max drafts, 480P, 5 s	12 takes	$3.00
K3 video reviews of drafts	12 reviews	~$0.50 (est.)
H3 Max finals, 768P, 10 s	3 shots × 2 takes	$4.80
Total		~$9.30

Rendering, not writing, dominates the bill, and bad prompts are what multiply renders.

Using K3’s vision to review the render

This is the step most guides skip. Per Moonshot’s vision docs, kimi-k3 accepts MP4, MOV, WebM and other video formats. You upload the file through the files API and reference it as a video_url part with an ms:// file ID (images go in as base64 or file IDs, not public URLs). Moonshot recommends video no larger than 1920x1080, which H3 Max’s 1344x768 output fits, and caps the request body at 100MB.

Keep the review prompt narrow:

Attached: the rendered clip and the JSON prompt that produced it.
Answer PASS / FAIL with a one-line reason for each:
1. Camera move type, amplitude and speed match each shot's spec.
2. Cuts land within 0.5 s of the specified timestamps.
3. Recurring subject matches the CHARACTER BIBLE.
4. On-screen text is legible and spelled exactly as quoted.
5. Hands, faces and product geometry show no visible warping.
6. Lip movement is plausible for any dialogue.
Then propose ONE prompt edit most likely to fix the worst FAIL.

One edit at a time keeps you from changing five variables and learning nothing. The limit: Moonshot documents K3’s video understanding for the picture but says nothing about analyzing the audio track, so check soundscape, music and voices by ear.

When this pairing makes sense

Your situation	Recommendation
Social clips, ad variants, storyboards	K3 + H3 Max: 480P drafts, 768P finals. The sweet spot
Multi-shot story with recurring characters	K3 with a cached bible; H3 Max per shot; cut in an editor
2K or broadcast delivery	K3 + regular H3 (2K at $0.13/s)
Heavy multi-reference work	Regular H3; confirm reference support before using Max
One-off clip, no pipeline	Skip K3; the Hailuo web app or fal playground with expansion on is simpler
You already code with Kimi	Script the loop with Kimi Code and K3
Strict data residency	Choose providers deliberately; see the K3 provider guide

Limitations and honest caveats

• K3 is slow for interactive work. At about 34 output tokens per second, a max-effort prompt can take longer to write than H3 Max takes to render. Lower the effort.
• Prompt expansion can undo your work. fal exposes prompt_expansion_mode (disabled, balanced, quality) and recommends balanced; with a precise K3 prompt, A/B test disabled on the same seed. MiniMax’s own docs describe no expansion switch.
• Provider specs disagree. MiniMax lists Max at 480P/768P, fal adds 1080P; MiniMax says Max takes references, Vercel says text or a starting image. Read your provider’s schema.
• Benchmarks are early and vendor-reported. They say nothing about your faces, products or text overlays.
• K3 cannot fix a vague brief. A generic idea gets a competent, generic shot.
• Fifteen seconds is the ceiling. Longer pieces mean multiple clips, and continuity becomes your job.
• Rights and policy. H3’s license adds steps for US, EU, UK and South Korean users, hosted endpoints apply safety filters, and you should not prompt for real people or trademarked characters.

FAQ

Can Kimi K3 generate video by itself? No. Kimi K3 takes video as input but only outputs text. It can write shot lists, camera directions, dialogue and structured prompts, and it can critique a rendered clip, but the MP4 itself has to come from a video model such as MiniMax H3 Max.

What is MiniMax H3 Max? A post-trained, speed-optimized version of MiniMax’s open-weight H3 video model, built by fal and announced August 27, 2026. It renders 5 to 15 second clips at 480P or 768P (fal also lists 1080P), 24 FPS, with audio generated in the same pass. It is available as MiniMax-H3-Max on MiniMax’s API and through fal, Vercel AI Gateway, OpenRouter and Runware.

How much does one H3 Max clip cost? On MiniMax’s pay-as-you-go card, $0.05 per output second at 480P and $0.08 at 768P. A 5-second 480P draft is $0.25, a 10-second 768P clip is $0.80 and a 15-second 768P clip is $1.20. The Kimi K3 prompt-writing cost is usually a few cents per shot.

What prompt format does MiniMax H3 expect? MiniMax’s official prompt guide defines three parts: integrated_multimodal_description (visuals, action, dialogue and on-screen sound in time order), overall_soundscape and non_diegetic_music. Camera moves should state type, amplitude and speed. MiniMax’s API caps prompts at 7,000 characters.

Should I turn off prompt expansion when using Kimi K3 prompts? Test both. fal’s H3 Max endpoint offers prompt_expansion_mode (disabled, balanced, quality) and recommends balanced. Expansion can rewrite details K3 chose on purpose, so render the same prompt and seed both ways and keep whichever follows your shot spec better.

Can Kimi K3 review the videos H3 Max renders? Yes, for the picture. Upload the MP4 through Moonshot’s files API and pass it to kimi-k3 as a video_url part with an ms:// file ID. H3 Max’s 768P output is within Moonshot’s recommended 1920x1080 limit. Audio-track analysis is not documented, so check the sound yourself.

Should I use H3 Max or the regular MiniMax H3? H3 Max for fast, cheap iteration at 480P or 768P. Regular H3 when you need 2K ($0.13 per second on MiniMax), 4-second clips or heavy multi-reference inputs. At 768P both cost $0.08 per second, so speed is the reason to pick Max there.

The bottom line

Split the job: Kimi K3 decides what the shot is, H3 Max renders it. MiniMax’s guide gives a precise target format of three sections, timed cuts, speaker tags and camera moves with type, amplitude and speed. K3 hits that format consistently across a project, especially with a cached bible in its 1M-token context. With 480P drafts at $0.25 that render in seconds, you can afford a real review loop where K3 checks every take against the spec. The weak spots are predictable: provider specs that disagree, vendor-reported benchmarks, a 15-second ceiling and audio you still judge by ear.

Keep reading: the Kimi K3 vision guide, full K3 specifications, every K3 API provider compared, Kimi Code setup with K3 and the complete Kimi K3 guide.