TokenLX — One API for 245+ AI Models

MiniMax-H3 is a general-purpose omni-modal video model that understands text, image, video, and audio context together, and generates video with native stereo audio up to 15s at 2K. It supports text-to-video, first/last-frame I2V, and multimodal reference-to-video.

Key strengths

Omni multimodal context
Native stereo audio video
Up to 15s / 2K output
Reference image / video / audio control
Strong price-performance on 2K

Use cases

Cinematic short clips
Product & character videos
Multimodal reference editing
Storyboard-to-video drafts
Localized creative video

videoomni2kmultimodalaudio

MiniMax's minimax/MiniMax-H3 is a high-fidelity video generation model. It supports text-to-video and image-to-video workflows with configurable duration, aspect ratio, and resolution, plus first-frame and last-frame control for guided scene composition.

Generates cinematic clips with consistent motion, camera control, and optional native audio. Billed per second of generated content.

minimax/MiniMax-H3 is fully OpenAI-compatible — drop in your existing OpenAI Python or Node SDK and switch `baseURL` to `https://api.tokenlx.ai`. TokenLX transparently routes your requests to the optimal provider endpoint while preserving streaming, function-calling, and structured-output semantics.

Performance

Compare different providers across TokenLX · All locations.

Throughput

tok/s

Latency

E2E Latency

132

Tool Call Errors

0.05

Output Errors

0.36

Time to First Token

Effective Pricing

Pricing is shown by the model billing method, using per-call or per-second prices and resolution tiers.

Per second

$0.07

Per duration

Resolution pricing · with audio

$0.12

768P

$0.07

Recent activity

Total usage per day on TokenLX (last 30 days).

Prompt

132.00M

Completion

244.20M

30d ago15d agotoday

Sample code & API

TokenLX normalizes requests and responses across providers. Use any OpenAI SDK or our native SDK.

# ─── 文生视频示例 ───
import requests, time

headers = {"Authorization": "Bearer sk-aihub-...", "Content-Type": "application/json"}

resp = requests.post(
    "https://api.tokenlx.ai/v1/aigc/video/tasks",
    headers=headers,
    json={
        "model": "MiniMax-H3",
        "prompt": "Epic cinematic teaser: a captain stands before a starship observation window as the fleet jumps to lightspeed",
        "duration": 5,
        "resolution": "2K",
        "aspectRatio": "16:9",
    },
)
task = resp.json()
task_id = task.get("taskId") or task.get("task_id") or task.get("id")

# 轮询结果
while True:
    result = requests.get(
        f"https://api.tokenlx.ai/v1/aigc/video/tasks/{'{'}task_id{'}'}?model=MiniMax-H3",
        headers=headers,
    ).json()
    print("status:", result.get("status"))
    if result.get("status") in ("completed", "succeeded", "done"):
        print(result)
        break
    time.sleep(10)

# ─── 图生视频（首帧驱动）───
i2v_resp = requests.post(
    "https://api.tokenlx.ai/v1/aigc/video/tasks",
    headers=headers,
    json={
        "model": "MiniMax-H3",
        "prompt": "画面中的人物缓缓转头微笑",
        "referenceImageUrls": ["https://example.com/first-frame.jpg"],
        "videoInputMode": "first_frame",
        "duration": 5,
        "resolution": "2K",
        "aspectRatio": "16:9",
    },
)
print("taskId:", i2v_resp.json())
# 轮询方式同上

# ─── 图生视频（首尾帧驱动）───
fl_resp = requests.post(
    "https://api.tokenlx.ai/v1/aigc/video/tasks",
    headers=headers,
    json={
        "model": "MiniMax-H3",
        "prompt": "平滑过渡",
        "referenceImageUrls": [
            "https://example.com/start.jpg",
            "https://example.com/end.jpg",
        ],
        "videoInputMode": "first_last_frame",
        "duration": 5,
        "resolution": "2K",
        "aspectRatio": "16:9",
    },
)
print("taskId:", fl_resp.json())

# ─── 参考视频生成示例 ───
ref_resp = requests.post(
    "https://api.tokenlx.ai/v1/aigc/video/tasks",
    headers=headers,
    json={
        "model": "MiniMax-H3",
        "prompt": "保持相同的运动风格，换成海边场景",
        "videoInputMode": "reference",
        "referenceVideoUrls": ["https://example.com/reference.mp4"],
        "duration": 5,
        "resolution": "2K",
        "aspectRatio": "adaptive",
    },
)
print("taskId:", ref_resp.json())

Replace sk-aihubrouter-… with your key from the dashboard.

Parameter Reference

Async Workflow

Create task

POST /v1/aigc/video/tasks

Poll result

GET /v1/aigc/video/tasks/{taskId}?model=MiniMax-H3

Recommended poll interval: 5–15sTask ID validity: 24h

Response is raw upstream JSON passthrough — fields vary by channel.

MiniMax-H3 uses the omni Video V2 API. Billed per output second by resolution (768P / 2K). Prompt ≤7000 characters. Mixed reference files ≤12 total.

Parameter	Type	Required	Description
model	string	Yes	Model name
prompt	string	Yes	Video description prompt
duration	int	Yes	Duration in seconds (integer 4–15) `456789101112131415`
resolution	string	No	Output resolution (default 768P; 2K available) `768P2K`
aspectRatio	string	No	Aspect ratio. Text-to-video requires an explicit ratio (not adaptive). `16:99:161:14:33:421:9adaptive`
referenceImageUrls	array	Image-to-video	Reference images (≤9). First/last frame or multimodal reference. Formats: JPG/JPEG/PNG/WebP/HEIC; ≤30MB each.
referenceVideoUrls	array	Ref-video	Reference videos (≤3 clips, each 2–15s, total ≤15s, ≤50MB each)
referenceAudioUrls	array	No	Reference audio (≤3 clips, each 2–15s; must pair with image or video)
videoInputMode	string	No	Input mode. Auto-inferred if omitted: 1 image → first_frame, 2 → first_last_frame, ≥3 or with video/audio → reference `first_framefirst_last_framereference`

modelstringYes

Model name

promptstringYes

Video description prompt

durationintYes

Duration in seconds (integer 4–15)

456789101112131415

resolutionstringNo

Output resolution (default 768P; 2K available)

768P2K

aspectRatiostringNo

Aspect ratio. Text-to-video requires an explicit ratio (not adaptive).

16:99:161:14:33:421:9adaptive

referenceImageUrlsarrayImage-to-video

Reference images (≤9). First/last frame or multimodal reference. Formats: JPG/JPEG/PNG/WebP/HEIC; ≤30MB each.

referenceVideoUrlsarrayRef-video

Reference videos (≤3 clips, each 2–15s, total ≤15s, ≤50MB each)

referenceAudioUrlsarrayNo

Reference audio (≤3 clips, each 2–15s; must pair with image or video)

videoInputModestringNo

Input mode. Auto-inferred if omitted: 1 image → first_frame, 2 → first_last_frame, ≥3 or with video/audio → reference

first_framefirst_last_framereference

Image-to-video Capabilities

✓First frame

✓First + last frame

✓Reference image

Sample code & API

TokenLX normalizes requests and responses across providers. Use any OpenAI SDK or our native SDK.

# ─── 文生视频示例 ───
import requests, time

headers = {"Authorization": "Bearer sk-aihub-...", "Content-Type": "application/json"}

resp = requests.post(
    "https://api.tokenlx.ai/v1/aigc/video/tasks",
    headers=headers,
    json={
        "model": "MiniMax-H3",
        "prompt": "Epic cinematic teaser: a captain stands before a starship observation window as the fleet jumps to lightspeed",
        "duration": 5,
        "resolution": "2K",
        "aspectRatio": "16:9",
    },
)
task = resp.json()
task_id = task.get("taskId") or task.get("task_id") or task.get("id")

# 轮询结果
while True:
    result = requests.get(
        f"https://api.tokenlx.ai/v1/aigc/video/tasks/{'{'}task_id{'}'}?model=MiniMax-H3",
        headers=headers,
    ).json()
    print("status:", result.get("status"))
    if result.get("status") in ("completed", "succeeded", "done"):
        print(result)
        break
    time.sleep(10)

# ─── 图生视频（首帧驱动）───
i2v_resp = requests.post(
    "https://api.tokenlx.ai/v1/aigc/video/tasks",
    headers=headers,
    json={
        "model": "MiniMax-H3",
        "prompt": "画面中的人物缓缓转头微笑",
        "referenceImageUrls": ["https://example.com/first-frame.jpg"],
        "videoInputMode": "first_frame",
        "duration": 5,
        "resolution": "2K",
        "aspectRatio": "16:9",
    },
)
print("taskId:", i2v_resp.json())
# 轮询方式同上

# ─── 图生视频（首尾帧驱动）───
fl_resp = requests.post(
    "https://api.tokenlx.ai/v1/aigc/video/tasks",
    headers=headers,
    json={
        "model": "MiniMax-H3",
        "prompt": "平滑过渡",
        "referenceImageUrls": [
            "https://example.com/start.jpg",
            "https://example.com/end.jpg",
        ],
        "videoInputMode": "first_last_frame",
        "duration": 5,
        "resolution": "2K",
        "aspectRatio": "16:9",
    },
)
print("taskId:", fl_resp.json())

# ─── 参考视频生成示例 ───
ref_resp = requests.post(
    "https://api.tokenlx.ai/v1/aigc/video/tasks",
    headers=headers,
    json={
        "model": "MiniMax-H3",
        "prompt": "保持相同的运动风格，换成海边场景",
        "videoInputMode": "reference",
        "referenceVideoUrls": ["https://example.com/reference.mp4"],
        "duration": 5,
        "resolution": "2K",
        "aspectRatio": "adaptive",
    },
)
print("taskId:", ref_resp.json())

Replace sk-aihubrouter-… with your key from the dashboard.

Parameter Reference

Async Workflow

Create task

POST /v1/aigc/video/tasks

Poll result

GET /v1/aigc/video/tasks/{taskId}?model=MiniMax-H3

Recommended poll interval: 5–15sTask ID validity: 24h

Response is raw upstream JSON passthrough — fields vary by channel.

MiniMax-H3 uses the omni Video V2 API. Billed per output second by resolution (768P / 2K). Prompt ≤7000 characters. Mixed reference files ≤12 total.

Parameter	Type	Required	Description
model	string	Yes	Model name
prompt	string	Yes	Video description prompt
duration	int	Yes	Duration in seconds (integer 4–15) `456789101112131415`
resolution	string	No	Output resolution (default 768P; 2K available) `768P2K`
aspectRatio	string	No	Aspect ratio. Text-to-video requires an explicit ratio (not adaptive). `16:99:161:14:33:421:9adaptive`
referenceImageUrls	array	Image-to-video	Reference images (≤9). First/last frame or multimodal reference. Formats: JPG/JPEG/PNG/WebP/HEIC; ≤30MB each.
referenceVideoUrls	array	Ref-video	Reference videos (≤3 clips, each 2–15s, total ≤15s, ≤50MB each)
referenceAudioUrls	array	No	Reference audio (≤3 clips, each 2–15s; must pair with image or video)
videoInputMode	string	No	Input mode. Auto-inferred if omitted: 1 image → first_frame, 2 → first_last_frame, ≥3 or with video/audio → reference `first_framefirst_last_framereference`

modelstringYes

Model name

promptstringYes

Video description prompt

durationintYes

Duration in seconds (integer 4–15)

456789101112131415

resolutionstringNo

Output resolution (default 768P; 2K available)

768P2K

aspectRatiostringNo

Aspect ratio. Text-to-video requires an explicit ratio (not adaptive).

16:99:161:14:33:421:9adaptive

referenceImageUrlsarrayImage-to-video

Reference images (≤9). First/last frame or multimodal reference. Formats: JPG/JPEG/PNG/WebP/HEIC; ≤30MB each.

referenceVideoUrlsarrayRef-video

Reference videos (≤3 clips, each 2–15s, total ≤15s, ≤50MB each)

referenceAudioUrlsarrayNo

Reference audio (≤3 clips, each 2–15s; must pair with image or video)

videoInputModestringNo

Input mode. Auto-inferred if omitted: 1 image → first_frame, 2 → first_last_frame, ≥3 or with video/audio → reference

first_framefirst_last_framereference

Image-to-video Capabilities

✓First frame

✓First + last frame

✓Reference image

minimax/MiniMax-H3

Key strengths

Use cases

Performance

Effective Pricing

Recent activity

Sample code & API

Parameter Reference

More models from MiniMax

minimax/MiniMax-H3

Key strengths

Use cases

Performance

Effective Pricing

Recent activity

Sample code & API

Parameter Reference

More models from MiniMax