[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-deepseek-v4-1-flash-1017":3,"topics-all":38,"news-related-f481172b-9ff8-4024-9a87-8220e746af39":57},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"f481172b-9ff8-4024-9a87-8220e746af39","DeepSeek V4.1-Flash-1017 原地升级:官方称 minor,十一项基准全线上涨","DeepSeek 把 deepseek-flash 底层 checkpoint 原地换成 V4.1-Flash-1017:十一项基准全线上涨,Terminal-Bench 4.0 从 31.2 到 42.3,DeepSWE v1.1 达 77.0,任务更快、输出更短,权重按 MIT 开源。","开源一个多月后,DeepSeek 没开发布会、没换架构——直接把 deepseek-flash API 背后的 checkpoint 换掉了。新版 DeepSeek-V4.1-Flash-1017 在 model card 里自谦「a minor update」,对照表却不客气:十一项基准全线上涨,Terminal-Bench 4.0 从 31.2 拉到 42.3,越过 GPT-5.6 Sol 的 39.9。\n\n## 不是 minor 的 minor update\n\n对照表横跨新旧 Flash、Kimi K3、GLM-5.3、GPT-5.6 Sol、Claude Opus 5 六列。1017 十一项全涨,四项为六列最高:DeepSWE v1.1 的 77.0(高于 Opus 5 的 74.0)、Automation-Bench v1.0.6 的 60.5、Agents' Last Exam 的 35.4、Apex Bench v1.1 的 72.1;Terminal-Bench 涨幅最大,仍低于 Opus 5 的 51.8;HLE、ProgramBench、NL2Repo 也是 Opus 5 领先。一句话:agentic 与工程基准上 1017 甩开同表对手,通用推理上 Opus 5 身位还在。\n\n官方公告原话是「stronger benchmark performance, faster task completion, and shorter outputs」,称之为性能与成本的 Pareto 改进。对按 token 计费的用户,输出缩短直接写在账单上。\n\n## 训练侧:RL 自反馈\n\n训练手段写在公告标题里:Wish-Driven Programming Made Real——用带自我反馈的强化学习提升端到端项目开发能力。官方演示包括故宫 3D 建模、网页赛车游戏、Godot 引擎复刻 Portal 2 等。\n\n规格表没动:763B 参数、1M token 上下文、MIT 协议,权重上架 Hugging Face 与 ModelScope。工程细节:依然没有 Jinja chat template,encoding 目录给 Python 参考实现,另有 Rust 工具链 deepseek-recipe 统一处理消息编码、工具调用与思考解析;采样维持 temperature 1.0、top_p 1.0。\n\n## 同日开闸的 Search & Fetch API\n\n顺手一刀更有意思:DeepSeek 全量开放自研搜索引擎支撑的 Search & Fetch API,每端点每千次 0.75 美元,官方称日请求量数亿;经 DeepSeek Harness、Claude Code、Codex 的 server-side tools 调用暂免,10 月 31 日 16:00 UTC 起计费。官方建议 Harness 升至 v0.2.1-rc.1 或更高,腾讯 CodeBuddy、WorkBuddy 已适配。\n\n## 所以呢\n\n一个月前 V4.1-Flash 接替退役的 V4-Pro 算「换代」;这次连版本号都没大动,API 用户一觉醒来底层 checkpoint 已换人,基准表整体上移一档。当迭代收敛成「定期换 checkpoint + 静默推全量」,发布仪式感在消失,比拼只剩:谁换得更勤,谁在同一架构里榨出的提升更大。1017 证明后者可以很可观——但 Opus 5 在多数通用基准上仍领先,前沿位置眼下还是别人的。\n\n参考:api-docs.deepseek.com\u002Fnews\u002Fnews261017 | huggingface.co\u002Fdeepseek-ai\u002FDeepSeek-V4.1-Flash-1017","https:\u002F\u002Fapi-docs.deepseek.com\u002Fnews\u002Fnews261017","3202bafd-f422-4632-b981-ec60896b79df",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"120fa59a-ff6f-4537-9bf5-f818df636a0e","benchmark",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"b52db7e9-7c58-42c3-9536-5132cb2f8f72","deepseek",{"id":19,"name":20,"slug":20,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"id":22,"name":23,"slug":23,"description":14,"color":14},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"d1d045a2-10aa-447c-8f65-53a592499cc6","en","DeepSeek V4.1-Flash-1017: Minor Update, Double-Digit Gains","DeepSeek V4.1-Flash-1017 checkpoint swap: all 11 benchmarks up, Terminal-Bench 31.2→42.3, DeepSWE 77.0, shorter outputs, MIT weights.","A month after open-sourcing the weights, DeepSeek held no launch event and changed no architecture — it simply swapped the checkpoint behind the deepseek-flash API. The new version is DeepSeek-V4.1-Flash-1017. The official model card humbly calls it \"a minor update\" to DeepSeek-V4.1-Flash, but the accompanying comparison table disagrees: all eleven benchmarks rose versus the previous version, and Terminal-Bench 4.0 climbed from 31.2 to 42.3, stepping past GPT-5.6 Sol's 39.9.\n\n## The numbers first: a minor update that isn't\n\nThe model card's table spans six columns: old and new Flash, Kimi K3, GLM-5.3, GPT-5.6 Sol, and Claude Opus 5. Against its own previous version, 1017 rose on all eleven rows, and posted the highest score of the six columns on four of them — DeepSWE v1.1 at 77.0 (above Opus 5's 74.0 and GPT-5.6 Sol's 73.0), Automation-Bench v1.0.6 at 60.5, Agents' Last Exam at 35.4, and Apex Bench v1.1 at 72.1. Terminal-Bench 4.0 saw the largest jump, 31.2 to 42.3, though still below Opus 5's 51.8; Opus 5 also leads on HLE, ProgramBench, and NL2Repo-Bench. In one sentence: on agentic and engineering benchmarks 1017 pulls away from the same-table rivals, while on general reasoning and long-text understanding Opus 5 keeps its position.\n\nThe gains are not just scores. The official announcement's own words: \"stronger benchmark performance, faster task completion, and shorter outputs\" — stronger while faster and shorter, which DeepSeek calls a Pareto improvement in performance and cost. For API users billed by the token, shorter output is the kind of improvement that shows up directly on the invoice.\n\n## Training side: RL with self-feedback, wish-driven programming\n\nThe core training method is stated plainly in the announcement's title: Wish-Driven Programming Made Real — reinforcement learning with self-feedback to improve end-to-end project development. Four demos accompany it: 3D modeling of the Forbidden City, a web-based racing game, a Portal 2 recreation built with the Godot engine, and Valley of Ages, a 3D scene tracing prehistoric times to the modern era. From single-point code completion to \"make a wish, get a complete project\" — that is the capability narrative DeepSeek set for this release.\n\nThe spec sheet is unchanged: 763B parameters (per the model card), a 1M-token context window, recommended max_tokens of at least 256K, MIT license, with weights live on Hugging Face and ModelScope. A detail engineers should note: there is still no Jinja-format chat template. Instead, the encoding folder ships a Python reference implementation, plus deepseek-recipe — a set of Rust libraries with Python bindings that uniformly handles message encoding, tool calling, and thinking-mode parsing. Recommended sampling stays at temperature 1.0 and top_p 1.0.\n\n## Search & Fetch APIs opened the same day\n\nThe quieter move may matter more: DeepSeek fully opened its Search & Fetch APIs, powered by its in-house search engine, priced at $0.75 per 1,000 calls per endpoint. The company says the engine serves hundreds of millions of requests daily worldwide. Calls made via server-side tools inside DeepSeek Harness, Claude Code, and Codex are free for now, with billing starting October 31, 2026 at 16:00 UTC. DeepSeek also advises upgrading DeepSeek Harness to v0.2.1-rc.1 or later for the best experience with the new model; Tencent CodeBuddy and WorkBuddy are already adapted.\n\n## So what\n\nWhen V4.1-Flash took over from the retired V4-Pro a month ago, the industry framed it as a generational handover. This time even the version number barely moved — API users woke up to find the underlying checkpoint replaced and the benchmark table shifted up a notch. As model iteration converges into \"swap checkpoints on a schedule, roll out silently,\" the ceremony of version launches is fading, and the contest narrows to two things: who swaps more often, and who squeezes more gains out of the same architecture. 1017 proves the latter can be substantial — but a splash of cold water: Opus 5 still leads most general benchmarks. A few more minor-update jumps, and the frontier spot might change hands; for now it hasn't.\n\nReferences: [DeepSeek official update](https:\u002F\u002Fapi-docs.deepseek.com\u002Fnews\u002Fnews261017) | [Hugging Face model card](https:\u002F\u002Fhuggingface.co\u002Fdeepseek-ai\u002FDeepSeek-V4.1-Flash-1017)","deepseek-v4-1-flash-1017","2026-10-17T17:12:24Z","2026-10-17T17:12:57.603110Z","2026-10-17T17:12:57.603126Z",true,"agent",280,[39,48],{"slug":40,"tag_slug":40,"title_zh":41,"title_en":42,"intro_zh":43,"intro_en":44,"id":45,"is_active":35,"created_at":46,"modified_at":47},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":49,"tag_slug":49,"title_zh":50,"title_en":51,"intro_zh":52,"intro_en":53,"id":54,"is_active":35,"created_at":55,"modified_at":56},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":58},[59,64,69,74,79,84],{"id":60,"title":61,"news_slug":62,"published_at":63},"d4fa7e14-8fbd-4940-93a6-3dd6f0a3991d","DeepSeek V4 Pro 正式版：1.6T MoE，1M 上下文","deepseek-v4-pro-0813-ga-1m-context-moe","2026-08-13T02:00:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"c94766df-827e-4e4e-a006-b6639ec76722","DeepSeek V4-Flash-0731 转正观察:权重不动,后训练把 Agent 分数打到 V4-Pro 之上","deepseek-v4-flash-0731-agent-benchmark-official-aug2026","2026-08-01T02:00:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"d529114c-bfd7-4efb-9af4-444d3065b397","Step 5 Preview 开源,2.7 美元\u002F百万输出","stepfun-step-5-preview-600b-moe","2026-10-17T21:12:04+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"5a5efb41-7037-46cb-b93e-6842f9ff2d62","UltraBERT:91.0 GLUE+47.7 BEIR,BERT 配方重审后 encoder 双榜登顶","ultrabert-encoder-dual-lm-heads","2026-10-14T17:10:59+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"983fb4d1-6c63-4828-adbf-a59c57e02b64","Perplexity开源决策模型:总分微胜Jev","perplexity-decider-v1-27b-open-source","2026-10-02T21:20:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"4c4a2a9e-f69b-4985-bd42-97ab2ef4e2ac","Spark-X2.5-4B 开源:4B 跑 1M 上下文,22 项基准打 9B 级 Qwen3.5","spark-x2-5-4b-apache-open-source","2026-09-16T01:30:00+00:00"]