DeepSeek updates V4.1 Flash with 1017 refresh

DeepSeek released V4.1-Flash-1017 on October 17, calling it a minor update to September’s V4.1 Flash model with stronger reasoning, coding, agentic and vision performance.

The update is available on DeepSeek’s website, app and API. Developers can continue using the deepseek-flash model name, and DeepSeek said the calling method and pricing remain unchanged. The company also claims the model completes tasks faster and produces shorter outputs than its predecessor, though it did not quantify those reductions in the opened materials. The available evidence includes no independent reproduction of its benchmark, runtime or output-length results.

In DeepSeek’s benchmark table, V4.1-Flash-1017 scored 42.6 on HLE without tools, compared with 36.8 for V4.1 Flash. The company reported 42.3 versus 31.2 on Terminal-Bench 4.0, 77.0 versus 74.2 on DeepSWE v1.1 and 29.1 versus 20.3 on ProgramBench.

DeepSeek reported gains on agentic tests as well. On Automation-Bench v1.0.6, the new checkpoint scored 60.5 against 54.8 for the September model. Its claimed vision results included 83.0 versus 78.9 on tool-assisted Chartography and 92.5 versus 89.6 on tool-assisted BabyVision.

The company said its Terminal-Bench result used the Claude Code harness, while its DeepSWE result used Mini-SWE-V2. The 44.0 figure shown alongside the main HLE score covers only the benchmark’s text-only subset.

DeepSeek said it used reinforcement learning with self-feedback to improve end-to-end project development, including game creation and 3D modeling. The release materials do not provide enough detail to isolate all training or architectural changes from the September checkpoint.

The September V4.1 Flash release introduced a multimodal mixture-of-experts model built on a 552-billion-parameter backbone and supporting contexts of up to 1 million tokens. For the 1017 refresh, DeepSeek published the model weights under the MIT license. The new model card recommends a 1-million-token context window.