Boost Agentic Performance with the Temporary V3.2-Speciale Model
Access the temporary high-compute endpoint for maximum reasoning performance.
Utilize the temporary 'Speciale' endpoint (available until Dec 15, 2025) for extreme reasoning tasks where standard models fall short.
The Scenario
You need to run a high-intensity reasoning benchmark or an agentic task that requires a model optimized for maximum compute and reinforcement learning.
Before & after
Researchers struggle with models that fail to follow complex agentic instructions, requiring them to manually fix 20-30% of the model's autonomous actions. This manual intervention can take half a workday.
Point your API calls to the 'V3.2-Speciale' temporary endpoint to leverage the scaled reinforcement learning for tough agentic workloads. This swap takes about 5 minutes to implement in your config file.
The Prompt
Using the V3.2-Speciale model, act as a high-performance agent to execute the following multi-step reasoning task: [DESCRIBE_AGENT_TASK]. Base your actions on maximum reasoning capability.
V3.2-Speciale is a high-compute variant available until December 15, 2025. It is optimized for maximum reasoning and agentic performance, reportedly outperforming GPT-5 on some reasoning workloads. Note: This specific endpoint does not support tool calls.
Source
Change Log | DeepSeek API Docs"DeepSeek-V3.2-Speciale is a high-compute variant... optimized for maximum reasoning and agentic performance. available until Dec 15th, 2025."