I’ve extensively tried the actual Deepseek R1 vs Sambanova’s. In every instance I’ve found that the “thinking” tokens are far lesser in Sambanova vs the actual Deepseek even though the output is well within the max tokens range. Can anyone shed some light why this is the case?
1 Like
@karanveers2020 I will have someone look into this.
-Coby
1 Like
Thank you, if required I can share examples over DM
Please do share examples. However it appears the primary difference is that our default system prompt in the playground which tells it to keep reasoning short.
As to using it over API Deepseek r1 recommends to avoid adding a system prompt; all instructions should be contained within the user prompt.
and example would be doing
response = client.chat.completions.create(
model="DeepSeek-R1",
messages=[{"role":"user","content":"""If a train leaves New York at 3 PM traveling at 60 miles per hour, and another train leaves Boston at 4 PM traveling at 80 miles per hour towards New York, when will they meet? Explain your reasoning step-by-step."""}],
Can you try that and see if your thinking response is within tolerance?
Coby