Docs / Account & Support
Provider Concurrency and Rate Limits
Sometimes you'll see "the upstream concurrency/rate quota for this model is full". This means an AI provider's quota for that API account is used up — your Soniva account is not being limited. This page lists each provider's limits and how Soniva handles them.
Sometimes you'll see "the upstream concurrency/rate quota for this model is full". This means an AI provider's quota for that API account is used up — your Soniva account is not being limited. This page lists each provider's limits and how Soniva handles them.
Figures checked on 2026-10-14. Providers change their quotas; the number shown in the provider's console wins. Items marked "to verify" are incomplete in public sources and should be confirmed in the console.
1. Why can millions use a provider's web chat while the API has limits?
- Web chat (the provider's own product) runs on the provider's whole fleet, scheduled by the provider.
- APIs get a per-account quota so one account can't take over shared capacity. Quotas usually rise automatically with identity verification and cumulative spend, or through a support ticket; most consoles have no "raise limit" button.
- Multiple API keys under one account share one quota — they don't add up.
2. AI chat / translation / AI bake-off
| Provider | Soniva default model | Standard-tier limit | Notes |
|---|---|---|---|
| DeepSeek | deepseek-v4-flash | 2,500 concurrent (v4-pro: 500) | Only concurrency is published |
| Alibaba Cloud Bailian | qwen3.7-flash | 30,000 requests/min, 5M tokens/min | Old snapshot versions go as low as 60/min |
| Volcano Engine Ark | doubao-seed-2-0-mini | 30,000 requests/min for the series | To verify: TPM for new models |
| Zhipu GLM | glm-4.7 | Free resource-pack tier: 1 request/sec; paid developer tier: 50/sec | Tier depends on verification and cumulative spend |
Zhipu's free resource-pack tier is the only one of the four that really maxes out after two or three parallel chats.
3. Text-to-speech
| Provider | Default limit | Notes |
|---|---|---|
| Doubao Speech (Volcano Engine) | 10 concurrent by default | More concurrency can be purchased |
| Alibaba Cloud CosyVoice | cosyvoice-v3-flash: 180 requests/min | To verify: other tiers |
| Tencent Cloud | Premium / large-model voices 20 concurrent; ultra-natural and one-sentence clone 10 | Can request more |
| Baidu AI Cloud | Paid 100/sec; free 5/sec | |
| MiniMax | Free users 10/min; paid users 20/min | Shared by main and sub-accounts |
| Microsoft Azure | Free F0: 20 per 60 sec (not adjustable); Standard S0: 30/sec default, up to 1,000 on request | F0 isn't meant for production |
| ElevenLabs | Flash/Turbo: Creator 10, Pro 20, Scale 30 concurrent; other models: 5, 10, 15 | By subscription tier |
| Cartesia | Free 2, Pro 3, Startup 5, Scale 15 concurrent | By subscription tier |
| OpenAI | gpt-4o-mini-tts: 10,000/min at Tier 1 | Rises with usage tier |
| Google Gemini TTS | Free tier 3/min, 15/day | Rises once billing is enabled; to verify: paid figures |
4. Speech-to-text (dictation)
| Provider | Default limit |
|---|---|
| Doubao Speech (Volcano Engine) streaming ASR | 10 concurrent by default, more can be purchased |
| Tencent Cloud | To verify |
5. How Soniva handles limits
- Server-side queue: everyone using Soniva credits shares Soniva's single account at each provider. The server caps how many requests run at once per provider; when full, requests wait in line (up to 25 seconds for chat and translation, 15 for text-to-speech, 10 for speech-to-text) instead of piling onto one provider. If a request still can't get a slot it doesn't just fail: you get an "upstream busy" message, a bake-off retries automatically, and in a normal chat you can simply resend. AI chat, translation, text-to-speech and speech-to-text all go through it.
- Automatic retry: when a provider says the quota is full, the server waits a second or two and retries twice; if it still fails, an AI bake-off run goes back to "waiting" and tries again after 5 and 10 seconds. "Stop" interrupts the wait.
- No endless waiting: if a model produces nothing at all (no answer and no reasoning) for two minutes, the request is stopped with a timeout message. The usual cause is a provider queue at peak times: the connection stays open but the model never starts answering.
- No credits charged for failures: if a request finally fails, its pre-charged credits are refunded.
- An extra layer for AI bake-off: one bake-off sends many requests, so on your computer each provider only runs a few at once (Zhipu gets 1), and the next starts as soon as one finishes.
- Use your own API key: your own account and quota, not shared with anyone and not part of the queue above.
6. Tips when using your own key
- Verify your identity and add a little balance first: many providers give an unpublished starter quota before that, far below the standard tiers above.
- Subscription providers (ElevenLabs, Cartesia) depend on your plan; enable overage once the plan quota is used up.
- To see your account's actual limit, check the provider console's rate-limit or quota page.
Sources
Was this article helpful?
Anonymous feedback is enabled when you allow analytics and the service is available. You can also report an issue by email.
The email includes this article. You can also copy the details and send them to soniva.cde@gmail.com