Agent workflows often make anywhere from one to ten model calls before completing a task. DeepSeek V4 Flash seems designed for this kind of use because it is fast and inexpensive, but deployments appear to have a default limit of only 20 requests per minute. In practice, I can get throttled after just a couple of agent steps. Why is the quota based on requests instead of tokens? Lightweight tool-call requests can use very few tokens, so a request-based limit gets exhausted quickly even when overall token usage is low.
3 Answers
I agree that the available open-weight options have been rough for agent use. Some alternatives are unreliable, and a few do not support tool calling properly at all. A model can be cheap and fast on paper but still be impractical if the endpoint cannot handle repeated tool interactions.
The limit probably reflects how much serving capacity is available for DeepSeek compared with larger, more established models. Some deployments may have higher limits, but they can still run into capacity errors when demand spikes. For agent workloads, another model may currently offer a more reliable experience despite being more expensive.
The real issue is the quota type. Even if the provider has limited capacity and needs a restriction, request-based throttling is a poor fit for agents. A workflow can spend most of its calls on tiny tool or control requests while using very few tokens, yet those calls count exactly the same as a large generation. A token-based limit, or separate quotas for generation and tool calls, would make more sense.
Exactly. I understand limiting access when capacity is tight, but counting a tiny tool-related request the same as a large completion makes the model difficult to use for the workload it was intended to handle.