DeepSeek releases V4.1-Flash-1017: update brings gains in reasoning, code and agent tasks with shorter outputs

DeepSeek-V4.1-Flash-1017 is a minor update to the V4.1-Flash model, released today (October 17, 2026) by the Chinese company. The official announcement came via a post on X (@deepseek_ai) and the API changelog. The open weights model (MIT license) is available on Hugging Face and ModelScope, with relevant gains in reasoning, coding, agentic and vision benchmarks, plus faster execution and shorter outputs.

According to the official DeepSeek API changelog, V4.1-Flash was updated to version 1017. Web, app and API already use the new model. The calling method and price remain the same: simply use the identifier deepseek-flash.

The company describes the update as a Pareto improvement in performance and cost. Compared with the previous version, the official numbers (maximum reasoning effort, with harness notes) include:

  • HLE (without tools): 42.6 (44.0 on the text-only subset)
  • Terminal-Bench 4.0: 42.3
  • DeepSWE v1.1: 77.0
  • ProgramBench: 29.1
  • NL2Repo-Bench: 68.2
  • Automation-Bench v1.0.6: 60.5
  • Agents' Last Exam: 35.4
  • Apex Bench v1.1: 72.1
  • Chartography (with tools): 83.0
  • BabyVision (with tools): 92.5
  • ZeroBench-main (with tools): 51.0

In addition to score gains, the model reduces execution time and output length. DeepSeek highlights specific optimizations for harnesses such as DeepSeek Harness, Tencent CodeBuddy and WorkBuddy. A demonstration shows the model building complete projects from scratch (including games and 3D modeling) with fewer interactions, thanks to reinforcement learning with self-feedback.

Open weights are available at:

The base architecture remains the same as V4.1-Flash (large-scale MoE with Causal Encoder-Decoder design, up to 1 million token context and native multimodal support). The update is incremental, focused on post-training and practical efficiency.

Sources

Transparency: This content was created, edited or reviewed with the assistance of artificial intelligence. Information was cross-checked with public posts on X and sources available on the internet. Consult the original sources to verify the full context.

By GeekikiBot