MicrosoftMicrosoftThank you for the update on DeepSeek R1's performance and pricing. While the blog post mentions improved performance, I haven't observed these improvements in practice, echoing the sentiment of the previous commenter regarding the East US region.
I'm particularly concerned about the 4,096 token context window for the DeepSeek-r1 model in Azure AI Foundry. For a model marketed for reasoning capabilities, this limit seems surprisingly restrictive. When testing with complex reasoning tasks, such as evaluating intricate mathematical integrals like:
Evaluate the following integral: \[\int_0^{\pi} \max\left(|2\sin(x)|, |2 \cos(2x) - 1| \right)^2 \cdot
\min\left(|\sin(2x)|, |\cos(3x)| \right)^2 \, dx\]I've encountered truncation of the model's reasoning process, indicated by the output being cut off mid-thought within the model's thinking tags. This suggests the context limit is being reached prematurely and hindering the model's ability to fully process and respond to complex queries.
It's worth noting that other providers offer significantly larger context windows (e.g., 32,768 tokens on AWS for same model), which is often crucial for effective reasoning tasks.
Since the performance update announcement on February 26th, 2025, I have tested DeepSeek-r1 daily. Unfortunately, the current performance and context limitations significantly impact its utility for my use cases, particularly those requiring in-depth reasoning.
I hope this feedback is helpful as you continue to optimize DeepSeek-r1 on Azure AI.