火山方舟已上线 DeepSeek-V4.1-Flash-1017。 相较 9 月 10 日首发版本，新模型 Agent 与代码能力上全面跃升，在Terminal-Bench 4.0、DeepSWE v1.1、ProgramBench 等多项基准较上版显著提升；同时任务运行时间与输出长度明显缩短，在多个主流 Harness 上达到新的性能–成本帕累托最优；基于自反馈的强化学习，新模型在游戏开发、3D 建模等"从零构建完整项目"的"许愿式"编程任务上能力显著增强。
在标准版基础上，火山方舟同步推出 DeepSeek-V4.1-Flash 经济版，闲忙时均享 8 折。以双版本供给满足企业在不同业务场景下对性能、时延与成本的差异化需求。无论是面向实时交互的智能应用，还是大规模后台及批处理任务，企业都能以更具性价比的方式调用 DeepSeek-V4.1-Flash，进一步加速 AI 应用落地。
标准版：官方效果一致，方舟能力加码
火山方舟 DeepSeek-V4.1-Flash-1017 与官方版本保持一致，企业无需自行部署和维护复杂架构，即可零门槛调用“满血版”模型能力。
在原厂能力基础上，火山方舟还提供视频理解、显式缓存、JSON Schema、前缀续写、role=developer 等增强特性，并可无缝衔接火山引擎云与 AI 生态，灵活调用联网搜索、MCP 等多种工具，帮助企业更高效地构建智能化、自动化应用。
在火山方舟体验： szacq.cn/juGUn/
经济版：效果完全一致，闲忙时均享 8 折
在标准版基础上，火山方舟同步推出 DeepSeek-V4.1-Flash 经济版。
在火山方舟体验：szacq.cn/jhR8D/
经济版与标准版使用同一份模型权重，Agent、代码、推理、多模态及 1M 上下文等能力完全一致，模型效果零损失；价格方面，经济版按标准版同时段刊例价的 8 折分时计费，任何时段调用都比标准版省 20%，以更低成本支持更大规模的模型调用。
除响应延时相对标准版有所增加外，经济版在效果与能力上均保持一致，尤其适合以下场景：
- 批量数据处理
- 后台 Agent 执行
- 训练数据合成
- 模型评测跑批
- 长文档理解
现有业务只需替换 Model ID（deepseek-v4-1-flash-economy-261017），即可快速完成迁移。
双版本灵活选择，让好模型用得起、用得好
标准版聚焦更优时延与实时体验，经济版以效果无损和闲忙时均享 8 折显著降低大规模调用成本。通过双版本组合，火山方舟让企业能够根据业务优先级灵活配置模型资源，在保持 DeepSeek-V4.1-Flash 完整能力的同时，更好地平衡性能、时延与成本，支持更多 AI 应用从验证走向规模化落地。
Agent Plan 同步升级 1017 版本
火山方舟 Agent Plan 个人版和企业版 DeepSeek-V4.1-flash 均升级至 1017 版本，同样提供了标准版和经济版供开发者选择。
Agent Plan 个人版的限时用量翻倍活动进行中，活动期间，不论标准版还是经济版，DeepSeek-V4.1-Flash-1017 抵扣系数再打 5 折，用量翻倍不加价，活动截止 12.30。
火山方舟是火山引擎旗下一站式大模型服务平台，以模型为核心，凭借专业算法服务、高并发算力保障、开发者友好体验与全周期安全可信四大支撑，为规模化模型服务落地保驾护航。平台已服务 110 万+ 企业与个人、覆盖 100+ 行业，让大模型深入千行百业。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
The volcano ark is on line DeepSeek-V4.1-Flash-1017. Compared to September 10, when it was launched, the new model Agent and code capabilities jumped across the board, and many of the benchmarks, such as Terminal-Bench 4.0, DeepSWE v. 1.1 and ProgramBech, were significantly higher than the previous edition; while the running time and output length of the task were significantly shorter, achieving the best of new performance-cost Parettos in a number of mainstream Harnesses; based on self-responding intensive learning, new models were significantly enhanced in game development, 3D modelling, etc., "Put-in-the-Wide" programming tasks "from zero complete project".
On a standard basis, the volcano ark is synchronized with the DeepSeek-V4.1-Flash Economic Edition, with an average of eight discounts for free time. The two-version supply meets the differentiated needs of enterprises for performance, latency and cost in different business scenarios. Whether it is a real-time interactive smart application, or a large backstage and batching mission, enterprises can further accelerate the landing of AI applications by calling DeepSeek-V4.1-Flash in a more cost-effective way.
Standard version: Officially consistent, ark capacity added
The Volcanic Ark DeepSeek-V4.1-Flash-1017, consistent with the official version, allows enterprises to deploy a “blooded version” model capability at the zero threshold without having to deploy and maintain complex structures on their own.
Based on the original plant capacity, the volcano ark also provides enhanced features such as video understanding, visible cache, JSON Schema, prefixes, rolle=developer, and seamless links to Volcengine clouds and AI ecology, flexible access to a variety of tools, such as network search and MCP, to help enterprises build more efficient intelligent and automated applications.
In the volcano ark experience: szacq.cn/juGun/
Economic version: full consistency, eight discounts for free time
On the basis of the standard version, the volcano ark simultaneously launched the DeepSeek-V4.1-Flash economic version.
Volcanic ark experience: szacq.cn/jhR8D/
The economic version is the same model weight as the standard version, which is fully consistent with the capabilities of Agent, code, reasoning, polymorphics and 1M context, with zero loss of model effects; on the price side, the economic version is based on the eight-center hourly rate for the same period of the standard version, and any time transfer is 20 per cent less than the standard version to support larger model calls at a lower cost.
With the exception of an increase in the response to the delay relative to the standard version, the economic version is consistent in terms of effectiveness and capacity and is particularly appropriate for the following scenarios:
- Batch data processing
- Backstage, Agent, execute.
- Training in data synthesis
- Model evaluation.
- Long document understanding.
The current operation requires only the replacement of Model ID (deepseek-v4-1-flash-economy-261017) to allow for rapid migration.
Two versions of flexible choices to make good models affordable and useful Okay.
The standard version is better focused on latency and real-time experience, with the economic version enjoying an average of 8 per cent of the cost of large-scale calls without loss of impact and free time. Through a two-version combination, the volcano ark allows enterprises to deploy model resources flexibly according to business priorities, while maintaining the full capacity of DeepSeek-V4.1-Flash, to better balance performance, latency and cost, and to support more AI applications from validation to size.
Agent Plan Sync Upgrade Version 1017
The Volcanic Ark Agent Plan personal and enterprise versions of DeepSeek-V4.1-Flash have been upgraded to version 1017, providing the same standard and economic versions for developers to choose.
The limited-time, a personal version of Agent Plan, is being used for a doubling exercise, during which the DeepSeek-V4.1-Flash-1017 credit factor is set at a further five times, doubling the volume to no price and ending the activity at 12.30.
The volcano ark is a all-in-one LLM service platform under the Volcengine flag, with models at its core, protected by four pillars of professional algorithm service, high-computability security, developer-friendly experience and full-cycle security, to serve as escorts to scale models. The platform has served 1.1 million + businesses and individuals, covering 100 + industries, allowing LLM to go into a wide range of industries.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.