WaveSpeedAI API alternative: what Sume offers instead

Is Sume an alternative to WaveSpeedAI?

For production media generation, yes, with a trade-off in breadth. WaveSpeedAI describes itself as unified API access to 1,000+ models for text-to-image, image-to-video, text-to-video and audio generation. Sume ships a much smaller managed catalog (Image, Video, Avatar, Music) and wraps it in jobs, balance reservation and signed webhooks. If breadth of models is the deciding factor, WaveSpeedAI wins.

What does WaveSpeedAI document?

The docs home lists models from Google, ByteDance, OpenAI, Stability AI, Luma and Runway, and names FLUX, Kling, Veo, Luma, Stable Diffusion, Seedance and Minimax among supported families. It has Submit Task and Get Result sections, plus pages on webhooks and streaming, so you can poll or receive a callback. The page I read did not spell out URL patterns, so look at the API reference before you write a client.

Rate limits are tiered by account level. The page lists Bronze as the default level at 5 predictions per minute and 2 concurrent tasks, then Silver (any successful single top-up below $1,000), Gold ($1,000 to $4,999) and Ultra ($5,000 or more) with higher per-minute and concurrency limits.

How does Sume handle bursts differently?

Sume separates dispatch from submit. If your workspace is at its processing concurrency, valid jobs are still accepted as queued, up to a queue capacity of max(3, concurrency x 5). Past that you get 429 queue_full. Ordinary request-rate limits return 429 rate_limited with retry-after, and a missing balance returns 402 insufficient_credits before any provider work starts.

The practical difference: on a tier model a top-up buys a higher ceiling; on Sume concurrency is set by plan, and the admission docs say prepaid top-ups do not raise it.

What about webhooks and results?

Sume webhooks are terminal-only: job.completed, job.failed and job.canceled, signed with HMAC SHA-256 over <timestamp>.<raw_body> in the x-sume-webhook-signature header. Keep a polling fallback anyway. Results come back as media.sume.com artifact URLs rather than provider URLs, so you do not depend on a third-party host expiring a link.

curl -X POST https://api.sume.com/v1/videos \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: ws-port-001" \
  -d '{"model":"sume/auto","prompt":"Slow dolly over a ramen counter at night"}'

How do you size a batch on each platform?

On a tiered account, your ceiling is a requests-per-minute number that follows your top-ups. The WaveSpeedAI page lists the default Bronze level at 5 predictions per minute, so a first batch is throttled until you move up. That shapes how you design a client: you pace submissions against the limit.

On Sume the shape is different. You can submit more jobs than your concurrency allows; they wait as queued. A Free workspace processes 1 job at a time and accepts 6 in total (1 processing plus 5 queued); Pro processes 4 and accepts 24; Scale processes 20 and accepts 120. The dashboard Concurrency tab shows the effective limit, exposed as generation_limits.concurrency_limit, and you should read that field rather than hard-coding the table.

This means a nightly batch on Sume is a matter of submitting up to the accepted capacity, then topping up as jobs finish. For a larger list of Format runs, a bulk queue accepts up to 100 items with a concurrency window, so your client does not need to drive the fan-out.

When should you stay on WaveSpeedAI?

Stay if you rely on a long-tail model Sume does not list, or if you want to pay down tiers by top-up. Move if you want one job envelope, idempotent retries and refunds recorded in a usage ledger (GET /v1/usage lists reservations, captures, refunds and top-ups). Read the admission rules before sizing a batch.