Article

SiliconFlow China rate limits: do usage tiers still help?

For teams using SiliconFlow China, check whether a usage-tier upgrade affects your model, find current limits and handle throttled requests.

Editorial illustration for SiliconFlow China rate limits: do usage tiers still help?: a geometric block represents a model release. Not documentary evidence.

SiliconFlow’s November 16, 2026 notice changes how much work some models on its China platform will accept over time. For teams running apps, agents or bulk jobs, usage tiers still matter for non-migrated paid models, but not for those moved to dynamic limits.

On migrated models, SiliconFlow says limits rise when load is light and tighten as load grows. L0–L5 entitlements stop applying even if they have not expired. The notice does not establish a token-price increase or the same policy for SiliconFlow’s international service.

Check your model before seeking an upgrade

Follow SiliconFlow’s documented checking procedure: sign in to Models, filter for dynamically limited models, open the model and read its Rate Limits section for current values and rules. The public notice does not list migrated models or specify a guaranteed minimum throughput.

The current rate-limit rules give three cases:

  • Dynamic model: plan around changing limits, not a tier upgrade.

  • Non-dynamic paid model: a higher usage level can raise its limits.

  • Non-dynamic free model: its limits remain fixed.

For eligible paid models, automatic upgrades depend on monthly consumption; the guide also offers tier packages through vendor contact. Confirm the model and commercial terms before paying.

Track requests and tokens across the account

RPM counts requests per minute; TPM counts tokens per minute—the text units a model processes. Either ceiling can constrain use. Limits apply per account, separately for each model, not per API key.

The operational implication: give a customer-facing app and background jobs using the same account and model a shared scheduler. Separate keys do not create extra capacity. Prioritise interactive work over deferrable processing.

Diagnose a 429 before spending more

SiliconFlow’s error reference identifies HTTP 429 as rate limiting; inspect its message for the specific metric. A 402 means account arrears, 401 concerns the API key, and 403 concerns permissions. These require different responses.

For 429s, SiliconFlow recommends pacing outgoing requests and exponential backoff—waiting progressively longer between retries. Set a retry budget appropriate to the task; that budget is an operational recommendation, not a documented vendor default.

Keep funding and capacity checks separate. The finance guide puts top-ups under Balance Recharge and usage records under Expense Bills. A funded account is not a promise of throughput, and recharging a balance is not the same as consuming it.

If your agent also retrieves web content, assess the tool service separately. The Firecrawl MCP account and usage limits guide covers that additional access decision.

When is another processing route worth considering?

Work that can wait may suit Batch Jobs. SiliconFlow says batch processing does not consume online rate limits, costs 50% less and typically finishes within 24 hours. That is a vendor expectation, not a guaranteed deadline; it does not suit a customer waiting for an immediate answer.

Confirm support for your chosen model: the batch documentation’s model list and code example differ. A job can expire with only some requests completed, and completed work remains billable. Inputs are capped at 1 GB and 5,000 lines; download results within 30 days.

If you are considering another provider, compare region, cost and model suitability before switching. The Qwen4-Max free API quota and spending-controls guide is another access checklist, not evidence of an equivalent replacement.