Models
The models behind Sume's video, image, speech, and music endpoints: ids, clip lengths, resolutions, aspect ratios, audio, references, and migration.
Start with What AI models does Sume support? Every model id, by API surface
Clef-omni reads video: a cheap check before a second video job
Cloudflare's Clef-omni takes audio and video input at $0.15 per million tokens. Sume does not offer Clef, so this is a pre-check pattern to run outside Sume.
FastH3 Trim: MiniMax H3 on an 8 GB GPU, and who may use it
FastH3 Trim prunes MiniMax H3 to 42 blocks and 8 steps. File sizes, the 8 GB claim, the quality trade, and what the H3 Community License allows.
Kandinsky 6.0 Video is MIT: can you use the clips commercially?
Kandinsky 6.0 Video Pro (29B) and Lite (3B) are MIT-licensed with joint audio. What MIT covers, what to check, and the hosted alternative on Sume.
Kimi K3 on Sume: no catalog row, and what to pick for a video agent
Kimi K3 is not in Sume's agent model catalog. Moonshot says it takes text, images and video; here is the verified alternative and how Kimi can still call Sume.
Korean clip? Whistle's seven languages vs Sume STT
Cactus Whistle lists English, German, French, Spanish, Italian, Dutch and Polish. For a Korean clip use Sume STT with a ko hint, then caption it.
Muse Spark 1.3 on Sume: picker row, tool use and media jobs
Meta says Muse Spark 1.3 uses about 20% fewer tool calls. Sume lists it as a catalog row behind the OpenRouter switch; the API cannot pick it.
Nova 2.5 Sonic for a narration file? Speech-to-speech vs Sume TTS
Nova 2.5 Sonic is built for live voice agents. For a finished narration file, Sume TTS takes a transcript up to 20,000 characters and returns audio as a job.
Qwen-Image-2.1-Turbo runs 8 steps; what does Sume's Qwen row expose?
Steps, CFG and seed are model-card settings. Sume's qwen/qwen-image row lists none of them: only ratio, n 1-4, references, output format. Check with one GET.
Reka Rho-1: can you generate video with it today?
Reka Rho-1 is a 19B research preview announced October 5, 2026, with no public weights, API or demo link. For video you can run today, check Sume's catalog.
Grok Imagine video ids: three at xAI, one on Sume
xAI lists grok-imagine-video, 1.5 and 1.5-lite; Sume's catalog has only grok-imagine-video-1.5. A quick map so you send the id that exists.
What is Grok Imagine Video 1.5 Lite, and what does Sume run?
Grok Imagine Video 1.5 Lite is xAI's low-price video model at $0.02 a second. Sume's catalog has grok-imagine-video-1.5 instead, image-to-video only.
Whistle 16.9 MB on-device model vs Sume STT: which clips go where
Cactus Whistle transcribes 30-second clips on the device; Sume STT bills $0.01 per audio minute in the cloud. Which audio belongs on which, with a table.
Whistle's 30-second window: passes for 5 minutes vs Sume STT
A 5-minute recording is ten 30-second Whistle passes on the device, or one Sume STT job at $0.05. How to plan chunking and when to skip it.
Whistle keyword biasing vs Sume STT: fixing product names in captions
Whistle can bias decoding toward keywords. Sume STT takes no vocabulary list, so for captions you pass script_text instead. How the two fixes compare.
MiMo V2.6 Pro and Flash in Sume's agent picker: what the catalog says
Sume's model catalog has rows for Xiaomi MiMo V2.6 Pro and Flash behind the OpenRouter switch. What the repo records, and what the API cannot pick.
Chinese AI video models by API: Kling, Wan, MiniMax limits and prices
Which Chinese video models can you call on Sume? kling-3, wan-3.0, minimax-h3 and minimax-h3-max: clip length, resolution, audio, references and price a second.
Eleven v4 on Sume lists 44 language codes, not 90+
ElevenLabs says v4 speaks 90+ languages. Sume's TTS Router catalog lists 44 ISO codes for eleven-v4, and 28 for Multilingual v2. Check yours before you send.
Gradium's 1,421 voices: how you pick a voice on Sume TTS
Gradium listed 1,421 TTS voices on Oct 7. Sume does not carry Gradium; here is how voice selection works on Sume TTS with avatar handles or voice ids.
Gradium's sub-50 ms TTS vs Sume's async TTS jobs: which fits
Gradium claims sub-50 ms latency. Sume TTS is a non-streaming job with a 30 second wait window; it suits batch narration and video, not live agent replies.
Which Sume video row takes 1:1, 3:4 or 21:9? Aspect ratio matrix
Seedance and MiniMax rows take 21:9; Wan 3.0 skips it; Kling 3 stops at 16:9, 9:16 and 1:1; Omni and Veo are 16:9 or 9:16 only. The full matrix by row.
Veo 3.1 4K needs an 8-second clip; Sume's Veo rows stop at 720p
Google offers Veo 3.1 1080p and 4K only at 8 seconds. Sume's Veo rows run 720p at 4, 6 or 8 seconds. See the gap and which Sume row reaches 4K.
Kling 3 on Sume: text or start and end frames, no reference inputs
Sume's kling-3 row takes text or start and end frames, 4 to 15 s at 1080p, with no reference images, video or audio. Prices and rows to use instead.
Mingmou 1 and Hunyuan 1.5 on Sume: fixed lengths and low rates
Sume's Tencent rows Mingmou 1 and Hunyuan 1.5 bill $0.05775 per second at 720p, take 5 or 10 seconds (Hunyuan 5 only), no audio. What they can and cannot do.
Gradium TTS covers 5 languages: what Sume's language field does
Gradium lists English, French, Spanish, German and Portuguese. Sume TTS takes a language field and a mismatch guard; here is how a non-English line behaves.
Aspect ratio auto on Sume: 6 image models list it, 8 edit models don't
Which Sume image models accept aspect_ratio auto when you edit a photo, which edit models reject it with a 400, and how to pick a ratio in code instead.
Shortest clip each Sume video row accepts, from 2 s Wan to 5 s H3
Minimum duration per Sume video row, from 2 seconds on Wan 3.0 to 5 on H3, with the billed price of that shortest clip at the cheapest resolution.
Seedream 5.0 Lite edit on Sume: no auto ratio, so set the shape
Of the three Seedream rows on Sume, only Seedream 4.0 lists an auto aspect ratio. For 5.0 Lite and 4.5 edits, pass an explicit ratio that matches your source.
Kling 3 on Sume: a 720p request renders 1080p at the same rate
Sume's Kling Video v3 Pro row has one resolution, 1080p. A 720p request is accepted and still renders 1080p. The rate depends on audio, not pixels.
Grok Imagine on Sume lists 9:19.5 and 9:20 but not 4:5
Sume's Grok Imagine row lists tall phone ratios 9:19.5 and 9:20 and wide 20:9 and 19.5:9, but not 4:5, 5:4 or 21:9, and caps n at 1. Here is what to do.
The 5 aspect ratios all 19 Sume image rows accept (a portable set)
All 19 Sume image rows list 1:1, 16:9, 9:16, 4:3 and 3:4. Use those five for a model bake-off and every row will accept the request without a 400.
Can Imagen 4 use my product photo? No: pick an edit-capable row
Imagen 4 Fast and Ultra on Sume are text-to-image only and reject references. Google also says Imagen is shut down in its API. Six rows take a product photo.
Veo 3.1 Lite (Preview) on Sume: what the preview label means
Sume's catalog names the row 'Veo 3.1 Lite (Preview)'. Its limits, the audio rates, and how it differs from the Gemini API's Lite lines for 720p and 1080p.
Higgsfield Soul on Sume: n must be 1 or 4, n=2 and n=3 return 400
Sume's Higgsfield Soul row accepts n of 1 or 4 only, rejects image_size, and offers 720p or 1080p at $0.005 or $0.0075 per image. Here is how to batch it.
Grok Imagine multi-image edit: xAI says 5 sources, Sume lists 10
xAI's Imagine docs cap multi-image edits at 5 source images and 10 outputs per request. Sume's Grok Imagine row lists 10 references and one output. Port safely.
Is Utopai X on Sume? It is built on MiniMax H3, which Sume lists
Utopai X runs inside Utopai's PAI platform and is post-trained on MiniMax H3. Sume does not list Utopai X; it lists minimax-h3 and minimax-h3-max.
Nano Banana 2.1 small text looks blurry? A three-row test on Sume
Google's Nano Banana 2.1 card says small text is often blurry. Run one poster prompt on three Sume rows for $5.40 for 12 prompts and pick by eye.
Nano Banana 2.1 replies with text and images; Sume returns URLs only
Google says Nano Banana 2.1 outputs text as well as images. Sume's image response lists image URLs and usage, so write captions and alt text in a separate step.
Veo 3.1 Lite limits: no 4K, no extension, and cheaper options
Veo 3.1 Lite is $0.05 a second at 720p but has no 4K and no extension, and it shuts down October 22. See what the Sume catalog offers in its place.
Veo 3.1 lists 4, 6 or 8 seconds; Omni on Sume runs 3 to 10
Veo 3.1 lists 4, 6 or 8 seconds. Omni on Sume runs 3 to 10 seconds at 1080p or 4K, priced here to the cent.
HappyHorse 1.1 image-to-video: 300 px, 20 MB and ratio 1:2.5 to 2.5:1
Alibaba's HappyHorse 1.1 image-to-video accepts JPEG, PNG or WEBP of at least 300 px, up to 20 MB. Prepare a first frame, or use an Omni image_url on Sume.
Hy Image 3.5 Preview: 17.44 s median latency vs Sume's 30 s wait
OpenRouter shows a 17.44 s median for Hy Image 3.5 Preview. Sume lists no Hy row, but its image calls wait 30 s, then return a 202 job. Handle both.
Hy Image 3.5 Preview at 91.57% availability: retry math for 200 images
OpenRouter shows 91.57% availability for Hy Image 3.5 Preview. For 200 images that is about 17 failed first tries. The math, and Sume's failed-job billing.
Seedream 5 Lite 4:5 portrait: render 1920x2400, resize to 1080x1350
Seedream 5 Lite lists 4:5 on Sume at $0.04375 per image. A 1920x2400 render shrinks to Instagram's 1080x1350 at an exact 0.5625 scale. 100 images cost $4.38.
Omni Flash 1.1 takes no audio reference: where a Short's voice goes
Gemini Omni Flash 1.1 accepts image and video references but not audio_url. For a voiced Short use native audio or a Timeline spine; 60 s is $7.60 at 720p.
H3 Max 4K request fails: H3 takes 2K and 4K as 768p upscales
minimax-h3-max accepts 480p, 768p or 1080p only. A 4K request belongs on minimax-h3, where it is a 768p upscale at $0.20 a second, $1.60 for 8 s on Sume.
Recast 480p fails: resolution must be 768p or 1080p, priced by length
H3 Max Recast refuses 480p with 'resolution must be 768p or 1080p'. A 10-second source costs $3.75 at 768p and $5.625 at 1080p on Sume. The 5 to 30 s table.
Omni Flash rejects 768p: the resolution strings each video model takes
Omni takes 360p, 720p, 1080p or 4K; H3 takes 768p, not 720p. One table of the accepted resolution strings per model and the price range of each on Sume.
How many jobs make a 3-minute vertical video: 6, 12 or 18 by model
A 180 s vertical video needs 6 jobs on Wan 3.0 or Seedance 2.5, 12 on MiniMax H3 and 18 on Omni Flash 1.1, since each model has a different maximum clip length.