Rate Limits

NexusFlow controls peak traffic via RPM, TPM, concurrent streams, async tasks, and a monitoring system. High concurrency is not a single number, but a combination of rate limiting, queuing, polling cadence, and model latency.

RPM

Controls request frequency, preventing instantaneous spikes from overwhelming upstream.

TPM

Limits tokens per minute, preventing long-context traffic from squeezing resources.

Concurrency

Long tasks should use the async queue rather than holding synchronous connections for extended periods.

Monitoring

Observe peak-period changes via TTFT, success rate, and per-model latency.

Plan Rate Limits

Plan	RPM	TPM	Concurrency	Description
Free	`20`	`40K`	`2`	Suitable for personal learning and testing
Developer	`60`	`150K`	`5`	Suitable for individual developers and small projects
Team	`200`	`500K`	`20`	Suitable for team collaboration and large-scale applications
Enterprise	`1000`	`2M`	`100`	Suitable for large-scale production environments
Custom	`Custom`	`Custom`	`Custom`	Custom limits based on your needs

Models Token Limitation

Models	Context window	Max Input	Max output
`qwen3-max`	262K	258K	64K
`qwen3.6-max-preview`	262K	262K	64K
`qwen3.6-plus`	1M	1M	64K
`qwen3.6-flash`	1M	1M	64K
`qwen3.5-plus`	1M	1M	64K
`deepseek-v4-pro`	1M	1M	16K
`deepseek-v4-flash`	1M	1M	16K
`deepseek-r1`	64K	64K	8K
`deepseek-v3`	64K	64K	8K

Rate Limiting Response Headers

Currently, the stable response headers you can rely on are those related to remaining balance. For more granular headers, refer to subsequent platform releases.

Response Header	Description
`X-RateLimit-Remaining`	Remaining request quota visible on the current request chain
`Retry-After`	Suggested wait time in seconds when rate limiting is triggered; clients should implement exponential backoff

High-Concurrency Scenario Recommendations

Sync vs. Async Traffic Separation

Chat goes to `/v1/chat/completions`, image/video goes to `/v1/tasks`. Separate long tasks from the synchronous path.

Smart Polling

Don't poll task status at high frequency; use a fixed 3-5 second interval or exponential backoff to reduce cascading amplification effects.

Monitor via the Dashboard

Track request volume, TTFT, success rate, and per-model latency changes to recognize whether you're approaching capacity limits.

Business-Side Degradation

During peak periods, prioritize switching to faster models, or downgrade by lowering max_tokens and long-context usage.