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
Google video model dates: Omni GA, Veo shutdowns, one table
One dated table of Google's video model lifecycle read from its own pages on Oct 3, 2026: Omni 1.1 Flash GA, Veo 3.1 preview shutdowns, and what to call.
GPT-6 Astra ultrafast tier vs Sume's Fast checkbox
OpenAI added service_tier ultrafast for GPT-6 Astra on Sep 29. Sume's picker has Fast but no ultrafast option.
Grok Imagine Image 2.0 policy review: cost of a refused image
xAI says generated media gets content policy review. Sume's docs say failed image generations are not billed; here is what is and is not documented.
Irodori-TTS-v4-Large: Japanese cloning, 120 s reference, terms
Irodori-TTS-v4-Large is a 3.29B Japanese TTS model with emoji style control and Gemma terms. What the card says and how Sume's audio tools fit.
Is FLUX.2 deprecated after FLUX 3? Keep production ids pinned on Sume
BFL's documentation says FLUX.2 remains fully supported for production image work. Pin the model id and read Sume's catalog before changing anything.
Kling 3.0 dialogue languages: is French or German covered?
Kling lists Chinese, English, Japanese, Korean and Spanish for VIDEO 3.0 dialogue. French and German are not on that list, so test a 720p clip on Sume first.
Kling 3.0 shot timing prompts: make seconds add up on Sume
Kling's guide writes multi-shot prompts as 'Shot 1, 4s'. On Sume kling-3 has no shot list field, so the timing lives in the prompt and in duration.
Kling 3.0 subject binding limits and what Sume kling-3 takes
Kling's subject binding takes up to 4 images, a 3-8 s clip and a voice sample. Sume's kling-3 takes none of them, so lock identity with a first frame.
Kling 3.0 in the Sume Videos panel: 5-15 s choices vs API 4-15 s
The panel offers Kling 3.0 at 5, 6, 8, 10 or 15 seconds and no end frame. The API accepts any whole second from 4 to 15 and a last frame.
LTX-2 diffusion decoder or convolutional decoder: which to use
LTX-2 ships a diffusion decoder (better quality, more VRAM) and a lighter convolutional one. What the README says, a draft-then-final habit, and hosted jobs.
LTX-2 Dub-It: lip-matched rephrasing vs Sume's dubbing steps
LTX-2's Dub-It pipeline rephrases speech while matching speaker and lips. What the README lists, what it omits, and what Sume's dubbing steps do and do not do.
Luma Ray 3.2 edit controls: pose, depth, normals, and Sume's edit
Luma's Ray 3.2 edit takes auto_controls, nine strength presets, or per-signal pose, depth, normals, trajectory and face controls. Sume's edit is prompt-only.
MiniMax-H3 Turbo LoRA: 4 to 8 steps, Apache 2.0, base licence
The MiniMax-H3 Turbo LoRA cuts sampling to 4 to 8 steps and lists Apache 2.0, but the 33B base keeps its own terms. What it needs and the hosted route.
MiniMax-Music3 weights: 5-minute songs, 8 GB, vs Sume Music Router
MiniMax-Music3 is open-weights: 5-minute 32 kHz songs, 8 GB with streaming, a $20M licence rule. What it needs, and what Sume's Music Router does instead.
Nano Banana Pro interleaved text and images vs Sume image output
Google documents Nano Banana Pro returning text blocks with illustrations in one answer. Sume's Image API documents image results only; here is the workaround.
OmniVoice: 600+ languages, CC-BY-NC weights, hosted TTS instead
OmniVoice covers 600+ languages in a 0.6B model, but its weights are CC-BY-NC. What the card says, what it omits, and where a hosted TTS job fits.
QuantFunc INT4 MiniMax H3: 3.2 s per step on an RTX 4090
QuantFunc's 4-bit MiniMax H3 claims 3.2 s per step on an RTX 4090. That is not a clip time. What the card says, what it omits, and when to use a hosted job.
Runway Veo negativePrompt: 1,000 characters, and Sume's options
Runway's API takes an optional negativePrompt of up to 1,000 characters on three Veo models. Sume's video request lists no such field, so use positive wording.
Seedance missing from the Sume Videos panel picker: what to do
The Videos panel picker lists Auto, Video 1.0, Kling 3.0, Wan 3.0, MiniMax and Grok. Seedance runs through the API by id, or through Auto.
Is Sume's kling-3 Kling VIDEO 3.0 or Omni 3.0? Read the catalog row
Sume's catalog names kling-3 'Kling Video v3 Pro' and lists no reference inputs. Kling describes Omni 3.0 as reference-driven, so match by capability.
Veda sparse attention for MiniMax H3: 6.8x attention, 3.1x clip
Veda's sparse attention keeps 10 percent of attention work for MiniMax H3. Why 6.8x on attention becomes 3.1x per clip, and what it needs to run.
Arabic text to speech API: send language ar to Sume TTS
Arabic is on Cartesia's Sonic 3.6, 3.5 and 3 but never on Sonic 2. How to send ar through Sume TTS, which voice id to use, and what an Arabic script costs.
Hebrew and Georgian text to speech API: he and ka on Sume TTS
Cartesia lists Hebrew (he) and Georgian (ka) on Sonic 3.6, 3.5 and 3. What to send to Sume TTS, why the library has no voice for them, and what a script costs.
Seedance 2.5 weak spots: physics and crowded scenes, a test plan
ByteDance says Seedance 2.5 still struggles with complex physics and multi-subject scenes. A 480p test plan with frame checks you can run on Sume.
Grok Imagine on Sume: 1080x1350 snaps to 2:3, 4:5 is a 400
On Sume, an exact size like 1080x1350 is snapped to Grok Imagine's nearest ratio (2:3), while aspect_ratio 4:5 is rejected. See which models keep 4:5.
Veo 3.1 Lite 1080p is 8 seconds only: options for a 5-second clip
Veo 3.1 Lite renders 1080p only at 8 seconds. To get a 5-second 1080p clip, trim an 8-second render with Sume or pick a model whose range includes 5.
Ideogram 4.5 output_format png returns 400 on Sume: omit it
On Sume, Ideogram 4.5 lists no output_format values, so even png is a 400. A table of which image models take png, jpeg or webp, and Recraft's webp-only.
Reference videos in AI video on Sume: Wan, H3, H3 Max and Omni limits
How many reference images, videos and audio files each Sume video model accepts: Wan 10/5/5, H3 9/3/3, Omni 10/3 with 3-second clips, Kling none.
Czech, Hungarian, Romanian, Greek TTS API: eight codes on Sume
Sonic 3.6 lists cs, sk, hu, ro, bg, hr, el and uk. Set language for Cyrillic and Greek text or the provider reads English. Sume TTS request, codes, and price.
Reference video or motion transfer: Seedance 2.5 or Genjutsu?
ByteDance says Seedance 2.5 reads a reference video beyond motion transfer. When to send reference_video_urls to Seedance and when to use Genjutsu on Sume.
Choose the music in a Gemini Omni clip by prompting the audio
Gemini Omni makes its own soundtrack. Google's guide shows how to steer it with a music style, a radio effect or a timed chorus; here is a cheap way to test.
Gemini Omni clip has sound I did not want: how to remove the audio
Gemini Omni always makes audio. Prompt for no dialogue, then drop the track with Sume video trim (audio drop) and confirm with video inspect.
MMAudio video-to-sound: CC-BY-NC weights and commercial use
MMAudio makes synchronized sound for video, but its checkpoints are CC-BY-NC 4.0 and default to 8 seconds. What that means for ads, and the Sume route.
claude-fable-5 is still active at Anthropic; Sume runs it as Fable 5.1
Anthropic keeps claude-fable-5 active until Jun 9, 2027, but a Sume Format run with that id runs on Fable 5.1. Vendor dates, price and the catalog alias.
Claude Opus 4.5-4.8 and Sonnet 4.6 retirement dates vs Sume ids
Anthropic lists retirement dates for Opus 4.5 through 4.8 and Sonnet 4.6. Sume's Format model field rejects those ids; the catalog row is Opus 5.5.
AI video looks fake? Check Seedance 2.5 frames for skin and light
ByteDance says Seedance 2.5 tuned textures, skin, eyes, lighting and stray subtitles. How to pull stills from a clip on Sume and check each one.
claude-haiku-4-5 is a 400 on Sume: the dotted id, and Haiku 5.5 news
Anthropic's alias claude-haiku-4-5 is not a Sume catalog id; Sume's row is claude-haiku-4.5 where the catalog lists it. Haiku 4.5 retires no sooner than Oct 15.
MiniMax H3 or H3 Max on Sume: which id for a 5 to 15 second clip?
Choose minimax-h3 for a 480p or 768p draft at $0.075 a second, minimax-h3-max when you need 1080p. Same 5-15 s range, same references; Max costs more at 768p.
Danish, Norwegian, Finnish text to speech API: da, no, fi on Sume
Sonic 3.6 lists da, no and fi. Norwegian is the code no, not nb or nn. How to send each through Sume TTS, which voice id works, and what a script costs.
How long can a Veo 3.1 video get? 148 seconds at 720p
Veo 3.1 extends clips 7 seconds at a time, up to 148 seconds at 720p. Sume has no extend task, so here are the longer single-clip models and how to join clips.
Which Sume image models accept quality? Only five do
Only five Sume image catalog rows list a quality field: GPT Image 2, 2.5 and Sunburst, Ideogram V3 and 4.5. The rest return 400 unsupported_parameter.
AudioCraft weights are CC-BY-NC: which models that covers
AudioCraft's code is MIT but its README puts the model weights under CC-BY-NC 4.0, including MusicGen and AudioGen. What that means for ad and video audio.
Veo 3.1 Lite takes text and images only: no references or extension
Google's Veo page shows Lite without reference images, video input or extension. If your Lite job needs those, here is what to use instead on Sume.
French text to speech API: dates, phone numbers and euros on Sume
Sonic 3.6 notes list fixes for French dates, times, phone numbers and Euro amounts. What reaches you through Sume TTS and what does not.
AI video signs and plates garbled? Write the text in the Omni prompt
Gemini Omni renders text well when you say what it reads. Google's guide covers signs, storefronts and plates; here is the prompt pattern and a Sume request.
Retirement notice: Anthropic 60 days, OpenAI 6 months. Pin the model?
Anthropic promises at least 60 days' notice and OpenAI at least 6 months for GA models. Whether to pin or omit the model field in a saved Sume Format.
Tamil, Telugu, Kannada, Malayalam TTS API: language codes on Sume
Sonic 3.6 lists bn, ta, te, kn, ml, mr, gu and pa beyond Hindi. How to send each on Sume TTS, what the Hinglish note does not cover, and the cost per script.
Stable Audio Open 1.0: 47-second limit and the Community License
Stable Audio Open 1.0 makes up to 47 seconds of 44.1 kHz stereo audio, trained on Freesound and FMA. Read the license note, limits, and a Sume alternative.