IT之家 10 月 17 日消息,DeepSeek V4.1 Flash 现已更新至 V4.1-Flash-1017 版本,官方网页端、APP 与 API 模型已同步变更为新模型(API 调用方式与售价保持不变,仍使用 deepseek-flash 即可访问最新模型)。
据深度求索官方介绍,此次更新在表现与成本之间达成了帕累托改进:在各项基准测试中,新版本的模型能力较上一版 V4.1 Flash 有显著提高,同时任务运行时间与输出长度均有缩短,使单次调用的综合成本下降。
DeepSeek 表示,新模型在提高表现的同时,还缩短了模型的任务运行时间与输出长度,让使用者可以以更低成本、更快速度完成任务,在多个 Harness 下达到新的表现-费用帕累托最优。
另外,DeepSeek 特别针对 DeepSeek Harness、腾讯 CodeBuddy、WorkBuddy 等平台做了适配优化,建议使用者将 DeepSeek Harness 升级至 0.2.1-rc.1 或更高版本以获取完整适配。
在能力维度上,V4.1-Flash-1017 引入了一项值得注意的变化:借助基于自反馈的强化学习训练,模型在从零构建完整项目维度的能力得到增强,覆盖游戏开发、3D 建模等情形。
官方称这一改进使模型可以承接更多“许愿式”编程请求,即在提示词和素材参考极为有限的条件下直接生成可运行的项目成果。官方演示案例表明,在 DeepSeek Harness 环境中,新版本以极短的交互轮次完成了完整项目的搭建。
同时,DeepSeek V4.1-Flash-1017 的模型权重已同步开放下载。DeepSeek 在 HuggingFace 和 ModelScope 上均给出了权重文件,延续了 V4.1 Flash 以 MIT 许可证开源的策略。
IT之家附地址:
本文来源:https://www.ithome.com/1/014/328.htm
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
IT House October 17th news, DeepSeek V4.1 Flash has been updated to V4.1-Flash-1017 version, the official web, app and API models have been synchronized to the new model (API call method and price remain unchanged, you can still use deepseek-flash to access the latest model).
According to the official introduction of Deep Search, this update has achieved a Pareto improvement between performance and cost: in various benchmarks, the model capability of the new version has been significantly improved compared with the previous version of V4.1 Flash. At the same time, the task run time and output length have been shortened, which reduces the comprehensive cost of a single call.
DeepSeek said that while improving performance, the new model also reduces the task runtime and output length of the model, allowing users to complete tasks at a lower cost and faster speed, and achieve new performance under multiple Harness - cost Pareto optimization.
In addition, DeepSeek has made adaptation optimizations for DeepSeek Harness, Tencent CodeBuddy, WorkBuddy and other platforms. It is recommended that users upgrade DeepSeek Harness to version 0.2.1-rc.1 or higher to obtain a complete adaptation.
In terms of capability dimensions, V4.1-Flash-1017 introduces a noteworthy change: with reinforcement learning training based on self-feedback, the model's ability to build a complete project dimension from scratch is enhanced, covering game development, 3D modeling, and other situations.
Officially, this improvement allows the model to take on more “wishful thinking” programming requests that directly generate actionable project outcomes with very limited prompt and material references. The official demo case shows that in the DeepSeek Harness environment, the new version completes the construction of the complete project in a very short interactive round.
At the same time, the model weights of DeepSeek V4.1-Flash-1017 have been synchronized and opened for download. DeepSeek gives weighting files on both HuggingFace and ModelScope, continuing the strategy of V4.1 Flash being open sourced under the mit license.
IT Home Address:
This article is from: https://www.ithome.com/1/014/328.htm For copyright issues, please write to: 13951431913#139.com or submit a review on the right.


