DeepSeek Updates Flash Model and Releases Open-Source Inference Framework

DeepSeek released an update to its V4.1 Flash model on October 17, reporting gains in coding and agentic tasks. The company claims the updated model, V4.1-Flash-1017, improved its Terminal-Bench 4.0 score from 31.2 to 42.3. Internal benchmark data indicates the model achieves a 77% success rate on the DeepSWE v1.1 coding test at a cost of $0.20 per task, positioning it against higher-cost competitors.

Following the model update, DeepSeek released DeepInfer, the inference framework currently used in its production services. Designed to manage computing resource allocation, the open-source framework includes support for Nvidia SM100 GPUs and Huawei Ascend 950 NPUs. The release intends to help developers optimize model latency and throughput through features like distributed inference and speculative decoding, though the documentation notes that some components, such as full load balancing, are not fully covered.

The company’s focus on efficiency coincides with broader market efforts to reduce the high costs of running large-scale AI agents. By providing tools to manage token consumption and hardware utilization, DeepSeek seeks to lower the barrier for API integration. However, the company’s stated benchmarks remain internal, and the practical effectiveness of the new framework will depend on how developers integrate the code into specific production environments.