Microsoft Azure Adds DeepSeek-V3.2 Models Focused on Enterprise AI Reasoning

Microsoft has expanded its Azure Foundry model catalog with the public preview release of DeepSeek-V3.2 and DeepSeek-V3.2-Speciale, two advanced reasoning models designed for enterprise and research workloads.

DeepSeek-V3.2 and DeepSeek-V3.2-Special are Now Available on Azure

DeepSeek-V3.2 targets long-horizon problem solving rather than simple text generation. The model focuses on agentic reasoning, multi-step decision making, and reliable tool usage across extended sessions. To achieve this, DeepSeek relies on DeepSeek Sparse Attention (DSA), a technique that filters unnecessary tokens in long contexts. This approach allows the model to reason faster while using significantly less memory, even at extreme context sizes.

According to DeepSeek, V3.2 delivers up to three times faster reasoning with similar output quality at 128,000-token context lengths, while dramatically reducing memory consumption. The model also dedicates a much larger portion of its training budget to reinforcement learning. More than 10% of total compute goes into reinforcement learning, compared to roughly 1% in traditional LLM pipelines. This shift improves multi-step reasoning, generalization, and consistency across complex tasks.

Another key addition involves a thinking retention mechanism. The model can preserve reasoning context across interactions, which increases efficiency and reliability during agent-based workflows. At scale, DeepSeek integrates agentic reinforcement learning to support long chains of thought, dependable tool use, and sustained reasoning across sessions.

For teams that prioritize raw reasoning power, DeepSeek-V3.2-Speciale takes a different approach. This variant emphasizes cognitive depth and extended chain-of-thought reasoning over structured outputs. Research groups, scientific teams, and evaluation labs benefit most from this model. DeepSeek reports frontier-level benchmark results, including best-in-class performance on Olympiad-style problems. To maximize reasoning capacity, the Speciale variant drops built-in functions and tool calling.

If you want to learn more, we suggest visiting Microsoft’s official blog post.

Both models are available today in public preview through Azure Foundry, with pay-per-token pricing listed directly in the model catalog.

While this is exciting news, Microsoft was a part of recently controversy, but the company denies overcharging UK businesses for using platforms other than Azure. What do you think about DeepSeek-V3.2 and DeepSeek-V3.2-Speciale joining Azure? Share your thoughts in the comments.