DeepSeek V3.2's Agentic Performance Surges 40% with Interleaved Thinking

DeepSeek V3.2's significant increase in agentic capabilities is attributed to a mechanism known as Interleaved Thinking, which has seen growing adoption within the open-source community. This approach addresses a common challenge in large language models: "state drift," where models lose track of initial instructions or long-term plans during extended interactions.

State drift occurs when models, despite initial logical processing, forget critical constraints or objectives after multiple rounds of conversation or complex tool interactions. For example, a model tasked with planning a family vacation might initially adhere to a "no strenuous activity" rule for an elderly participant but later suggest a demanding mountain hike after several itinerary modifications. This phenomenon is not due to a model becoming "stupid" but rather a limitation in how it processes and retains information over time.

Interleaved Thinking: A Solution to "Amnesia"

To counter state drift, major models, including Anthropic Claude, OpenAI GPT-OSS, MiniMax M2, and Kimi K2 Thinking, have adopted a similar technology, often referred to as "Thinking in Tool-Use" or "Interleaved Thinking." This method involves the model alternating between reasoning (thinking) and tool calling (action), continuously retaining and reusing the reasoning state from each round to achieve stable, cumulative long-term planning.

The predecessor to Interleaved Thinking, the early ReAct (Reasoning+Acting) paradigm, faced bottlenecks because its linear "observe -> think -> act" logic often simplified the thinking process. In implementations like OpenAI's Function Calling, this often meant the model directly outputted tool call instructions. When a tool executed and returned extensive data, the model struggled to integrate this new information without losing its original thought process, making it susceptible to environmental disturbances or irrelevant data. This "implicit reasoning" meant that the model's thought process was hidden and easily lost upon interruption.

Enhancing Robustness and Performance

Interleaved Thinking addresses this by making "thinking" an explicit, recorded state. Before each tool call, the model outputs a natural language segment containing its reasoning, often wrapped in tags like reasoning_details. This record serves as a memory for the model, guiding its subsequent actions and helping it maintain its long-term plan.

This mechanism has led to substantial performance improvements. For instance, on the SWE-Bench Verified leaderboard for software engineering tasks, enabling Interleaved Thinking improved MiniMax M2's performance by 3.3% (from 67.2 to 69.4). More dramatically, on web browsing tasks (BrowseComp), the improvement reached 40% (from 31.4 to 44.0), and on complex reasoning tasks like Tau², it improved by 36%.

The significant difference in performance across task types highlights Interleaved Thinking's role in resisting environmental disturbances. In low-disturbance environments like code, where error messages are clear, models can often recover. However, in high-disturbance environments like the internet, filled with noise such as advertisements and irrelevant content, traditional methods struggle. Interleaved Thinking acts as a filter, allowing the model to process and calibrate information explicitly, breaking down complex tasks into stable, atomic thinking loops.

Generalization Beyond Tools

The effectiveness of Interleaved Thinking also redefines the concept of agent generalization. While early industry thought focused on models learning to use more tools, MiniMax's team found that true generalization lies in adapting to disturbances throughout a task's trajectory. Different environments, prompt structures, and tool return formats can all disrupt a model's reasoning. Interleaved Thinking provides the model with self-correction capabilities by retaining reasoning content at each step, allowing it to align with the environment through explicit logical reasoning rather than relying on rote prompt templates.

Industry Adoption and Infrastructure Development

Despite its technical advantages, the implementation of Interleaved Thinking initially faced challenges due to lagging industry infrastructure. Most open-source tools were built on OpenAI's Chat Completion API, which did not natively support explicit "thinking processes." This led to users inadvertently discarding the reasoning_details field, hindering model performance.

In response, MiniMax actively engaged with the open-source community, submitting pull requests to major Agent development tools and platforms. This included modifying the underlying message processing logic in VS Code plugins like Cline to retain the model's thinking process in conversation history. Similar efforts were made for cloud IDEs like Kilo Code and model hosting platforms such as OpenRouter and Ollama, promoting the reasoning_details field as a de facto standard extension.

The inclusion of MiniMax M2 in the Amazon Bedrock model library, announced at AWS re:Invent 2025, further validates this approach. The recent releases of DeepSeek V3.2 and Kimi K2 Thinking, both incorporating Interleaved Thinking (or "Thinking in Tool-Use"), signal a growing industry consensus. Although specific API field names may vary (e.g., reasoning_details, reasoning_content, thinking_blocks), the underlying design philosophy emphasizes explicit, interleaved, and persistent thinking as crucial for the evolution of intelligent agents.

This shift reflects a broader understanding that AI performance scales not only with parameter count but also with test-time compute. Models are evolving from mechanical "copilots" to "thinkers" capable of pausing, reflecting, self-correcting, and executing long-chain tasks in complex, noisy real-world environments.