DeepSeek has released V4.1-Flash-1017, an update to its V4.1-Flash model that the company said improves reasoning, coding, agentic and visual tasks while reducing task completion time and output length.
The update is available across DeepSeek’s web product, app and developer platform. It supports a 1 million-token context window, image and text inputs, and is released under the MIT license.
According to DeepSeek’s technical report and model card, the model’s Terminal-Bench 4.0 score rose to 42.3 from 31.2 in the previous version. Apex Bench v1.1 also increased to 72.1 from 64.6, while coding and automation benchmarks posted similar gains.
The release builds on the V4.1-Flash architecture’s focus on KV cache compression, which reduces memory requirements during long-context inference. DeepSeek said the model can deliver stronger performance with shorter outputs, potentially improving latency and token efficiency for teams running agentic workloads at scale.
The company provides a Python reference encoder and a Rust-based toolkit for prompt formatting and response parsing. DeepSeek also said the model is being supported by its open-source community and can be deployed through additional infrastructure options.
Source: Tech in Asia
