DeepSeek API usage reports extremely inflated token counts and leaks DSML tool call markers

AI & Data Science NVIDIA NIM Models
1

model: deepseek-ai/deepseek-v4-flash

I’m using DeepSeek through the NVIDIA API endpoint, and I noticed that the usage field returned by the backend contains extremely inflated token counts compared with the actual character count.

Example response:

"usage": {
  "inputCharacters": 21864,
  "inputTokens": 878912,
  "outputCharacters": 84,
  "outputTokens": 1914,
  "totalTokens": 880826
}

The reported inputTokens value is about 40 tokens per input character, which seems unreasonable. With around 21k input characters, I would expect the token count to be much lower, not close to 879k tokens.

In addition, the model sometimes prints raw DeepSeek tool-calling markers in the response content, such as:

<｜DSML｜tool_calls

This makes me suspect that the DeepSeek tool-calling format may not be fully parsed or converted correctly by the NVIDIA compatibility layer. It may also be related to the inflated token usage if tool call chunks, hidden tool-call markup, or streaming chunks are being counted repeatedly.

Could you please help confirm:

  1. Is the usage.inputTokens value returned by the NVIDIA API expected to represent the actual billable token count?
  2. Is NVIDIA currently parsing DeepSeek DSML tool calls into OpenAI-compatible tool_calls fields?
  3. Could this be a bug in token accounting, especially when tools or streaming are enabled?
  4. Is there any recommended request format to avoid DSML markers leaking into content?

This issue makes it difficult to estimate cost and context usage accurately.

Thanks.

3

oh it doesn’t even work past 262144 Tokens now… that is if my 484000 token dataset and the subsequent error is to be believed.