Deepseek released V4.1 Flash 1017 on October 17, 2026, saying it excels in agentic tasks and coding, while significantly lowering per-task costs. It is the company's first major update to its Flash series since its initial release.
Deepseek reported a 77.0% benchmark score on the DeepSWE benchmark, measured at an average cost of just under $0.20 per task. That compares with its predecessor's 74.2% score at around $0.50 per run.
V4.1 Flash 1017 is built on reinforcement learning with self-feedback (RLSF) and targets end-to-end software development and agent tasks. Availability begins through the company's website, app, and API, initially for developers and researchers.
"Reinforcement learning with self-feedback lets the model use its own internal confidence scores or self-generated critiques as a reward signal," said Matthias Bastian, a contributor at THE DECODER. "This post-training method removes the need for human evaluators or external reward models."
The announcement follows Deepseek's expansion into open-source platforms like Hugging Face and ModelScope. Deepseek said the update represents a Pareto improvement in the tradeoff between performance and cost.
Deepseek did not say how the model will perform in independent benchmarks, and it raised the question of how developer impressions on code quality compare with benchmark results. The company said it will continue to refine the model and its tools.
Source: thedecoder
