DeepSeek V4 Flash seems designed for agentic applications, where one task can easily trigger anywhere from 1 to 10 model calls. It is fast and inexpensive, but deployments appear to have a default limit of only 20 requests per minute, and I can get throttled after just a couple of agent steps. Why is the quota based on requests instead of tokens? Lightweight tool calls may use very few tokens but still consume an entire request allowance, which makes the limit especially restrictive for agents.
4 Answers
The other open-weight options I tried were not very usable for agents. A few had tool support disabled entirely, while others struggled with tool calling, so switching models is not always a straightforward workaround.
The request limit is probably tied to capacity. There may simply be fewer servers available for this model than for the major hosted models, so the provider uses a conservative RPM cap to prevent overload.
Some deployments advertise a higher RPM, but capacity errors can still make them unreliable. I have had better results with Luna for agentic workloads, although its output pricing is considerably higher, which matters because generated output is usually the biggest part of the bill.
That explanation makes sense for having a limit, but request-based throttling still feels poorly suited to agents. A handful of small tool calls can exhaust the quota even when token usage is low. A token-based limit, or separate quotas for model generation and tool-related calls, would be much more practical.
Exactly. I understand limiting access when capacity is tight; the frustrating part is that the quota counts every request equally, regardless of how little work that request actually represents.
That tradeoff is the problem: the alternative may be more available, but paying more for output largely defeats the cost advantage that made this model attractive in the first place.