[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$f3p4nwztra0lsh":3},{"slug":4,"category":5,"publishedAt":6,"titleZh":7,"titleEn":8,"summaryZh":9,"summaryEn":10,"models":11,"hardwares":18,"bodyZh":19,"bodyEn":20,"sourceName":21,"sourceUrl":22,"updatedAt":23},"deepseek-v4-1-flash-1017-2026-10-19","NEWS","2026-10-19T10:00:00.000Z","DeepSeek-V4.1-Flash-1017 发布：免费 MIT 更新，推理更快、输出更短","DeepSeek-V4.1-Flash-1017 ships: free MIT update, faster reasoning, shorter outputs","10 月 17 日 DeepSeek 上线 V4.1-Flash 的 1017 检查点（MIT）。官方基准全面抬升：HLE 36.8→42.6、Terminal-Bench 31.2→42.3、Apex 64.6→72.1；输出更短带来实际更快与更省。本地量化当日铺开，站内模型目录已收录（763B 总参保持目录口径、无单卡实测）。","DeepSeek shipped the 1017 checkpoint of V4.1-Flash on Oct 17 (MIT). Broad official gains: HLE 36.8→42.6, Terminal-Bench 31.2→42.3, Apex 64.6→72.1; shorter outputs mean faster wall-clock and lower cost. Local quants landed same-week; catalog row added on our site (763B total stays catalog-only, no single-GPU runs).",[12,15],{"id":13,"name":14},18,"DeepSeek-V4.1-Flash",{"id":16,"name":17},85,"DeepSeek-V4.1-Flash-1017",[],"# DeepSeek-V4.1-Flash-1017 发布：免费 MIT 更新，推理更快、输出更短\n\n10 月 17 日，DeepSeek 在 Hugging Face 上线 **DeepSeek-V4.1-Flash-1017**（MIT）——原版 DeepSeek-V4.1-Flash 的日期检查点刷新。架构与部署路径不变（763.2B 总参 \u002F 37B 激活 MoE、1M 上下文），变的是权重质量与输出行为。\n\n## 基准增幅（官方口径）\n\n| 基准 | 原版 V4.1-Flash | 1017 | Δ |\n|---|---|---|---|\n| HLE（w\u002Fo tools） | 36.8 | **42.6** | +5.8 |\n| HLE text-only（†） | 39.1 | 44.0 | +4.9 |\n| Terminal-Bench 4.0 | 31.2 | **42.3** | **+11.1** |\n| DeepSWE v1.1 | 74.2 | 77.0 | +2.8 |\n| ProgramBench | 20.3 | 29.1 | +8.8 |\n| NL2Repo-Bench | 65.4 | 68.2 | +2.8 |\n| Automation-Bench | 54.8 | 60.5 | +5.7 |\n| Agents' Last Exam | 31.8 | 35.4 | +3.6 |\n| Apex Bench | 64.6 | **72.1** | +7.5 |\n| Chartography | 78.9 | 83.0 | +4.1 |\n| BabyVision | 89.6 | 92.5 | +2.9 |\n\nagentic 场景涨幅最猛（Terminal-Bench +11.1、ProgramBench +8.8），通用推理紧随其后（HLE +5.8、Apex +7.5）。\n\n## 输出更短 = 实际更快、更省\n\n官方明确把「更短的输出」列为卖点：同任务生成更少 token，意味着更低的端到端时延与更低的 API 费用。再叠加基准分提升，1017 是一次「同样的钱办更多事」的免费更新。\n\n## 部署注意\n\n- **没有 Jinja chat template**：对话格式改由官方 deepseek-recipe（Rust 库）处理，社区推理栈需跟进适配\n- 上下文 1M，官方建议 max_tokens ≥ 256K\n- 本地量化在发布后两天内铺开：GGUF、mxfp4、EXL3、GSQ-agentic-strata 均已出现\n\n## 站内收录口径\n\n模型目录新增 **DeepSeek-V4.1-Flash-1017**（与原版共享 37B 激活参数口径，MIT → 开源）。763B 总参的容量门槛意味着单卡组合进不了实测——原版 DeepSeek-V4.1-Flash 同样保持「目录收录、零实测记录」口径，本批不破例。想本地跑动它，等待的是多卡或专家卸载方案，不是更大的单卡。\n\n## 对比坐标系\n\n官方对比列里同时摆了 K3、GLM-5.3、GPT-5.6 Sol 与 Opus 5；媒体侧的两个焦点：「免费、MIT 许可的更新」，以及短输出对推理芯片需求预期的冲击（发布当日芯片股承压）。","# DeepSeek-V4.1-Flash-1017 ships: free MIT update, faster reasoning, shorter outputs\n\nOn Oct 17, DeepSeek published **DeepSeek-V4.1-Flash-1017** on Hugging Face (MIT) — a dated checkpoint refresh of the original DeepSeek-V4.1-Flash. Architecture and deployment path stay the same (763.2B total \u002F 37B active MoE, 1M context); what changed is weight quality and output behavior.\n\n## Benchmark gains (official numbers)\n\n| Benchmark | V4.1-Flash | 1017 | Δ |\n|---|---|---|---|\n| HLE (w\u002Fo tools) | 36.8 | **42.6** | +5.8 |\n| HLE text-only (†) | 39.1 | 44.0 | +4.9 |\n| Terminal-Bench 4.0 | 31.2 | **42.3** | **+11.1** |\n| DeepSWE v1.1 | 74.2 | 77.0 | +2.8 |\n| ProgramBench | 20.3 | 29.1 | +8.8 |\n| NL2Repo-Bench | 65.4 | 68.2 | +2.8 |\n| Automation-Bench | 54.8 | 60.5 | +5.7 |\n| Agents Last Exam | 31.8 | 35.4 | +3.6 |\n| Apex Bench | 64.6 | **72.1** | +7.5 |\n| Chartography | 78.9 | 83.0 | +4.1 |\n| BabyVision | 89.6 | 92.5 | +2.9 |\n\nThe biggest jumps are agentic workloads (Terminal-Bench +11.1, ProgramBench +8.8), with general reasoning close behind (HLE +5.8, Apex +7.5).\n\n## Shorter outputs = faster and cheaper in practice\n\nDeepSeek explicitly markets shorter outputs: fewer generated tokens per task means lower wall-clock latency and lower API bills. Combined with the score gains, 1017 is a free update that does more for the same money.\n\n## Deployment notes\n\n- **No Jinja chat template**: conversation formatting moved to the official deepseek-recipe (Rust library); community inference stacks need to follow\n- 1M context, max_tokens officially recommended at 256K or above\n- Local quants spread within two days: GGUF, mxfp4, EXL3 and GSQ-agentic-strata are all out\n\n## How we logged it\n\nOur model catalog now lists **DeepSeek-V4.1-Flash-1017** (shares the 37B active-params convention with the original, MIT → open source). The 763B total keeps it out of single-GPU benchmark scope — the original V4.1-Flash has the same catalog-only, zero-records stance and this batch does not break it. Running it locally is waiting on multi-GPU or expert-offload setups, not on a bigger single card.\n\n## Context\n\nThe official comparison column also lists K3, GLM-5.3, GPT-5.6 Sol and Opus 5. Press coverage focused on two angles: the free, MIT-licensed update itself, and how shorter outputs pressure the inference-chip demand narrative (chip stocks slipped on release day).","araoai","https:\u002F\u002Faraoai.com\u002Fnews\u002Fdeepseek-v4-1-flash-1017-2026-10-19","2026-10-19T12:45:28.296Z"]