[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-plamo-3-0-prime-pfn-japanese-domestic":3,"news-related-1d80585e-c797-4aa1-ac68-ef87334d5d0c":39},{"id":4,"title":5,"summary":6,"content":6,"original_url":7,"source_id":8,"tags":9,"translations":26,"news_slug":32,"published_at":33,"created_at":34,"modified_at":35,"is_published":36,"publish_type":37,"image_url":13,"view_count":38},"1d80585e-c797-4aa1-ac68-ef87334d5d0c","PLaMo 3.0 Prime 正式发布：PFN 把「日语实战」做成日本国产 LLM 的差异化战场","6 月 22 日，Preferred Networks（PFN）正式发布从零训练的国产生成式大模型 PLaMo 3.0 Prime。它不卷「谁更聪明」，而是把战场拉到日语业务的成本与合规上。\n\n技术上看，Prime 同步上线两个版本：Reasoning 模型应对复杂推理、专业问答与代码生成；Non-reasoning 模型面向文档摘要、客服、信息抽取等低延迟场景，二者共享一套权重族，可通过 API 切换。上下文从 β 版的 64K 扩到 256K——PFN 用 YaRN 加持续预训练实现，超过 gpt-oss-120b 的 128K、Claude Haiku 4.5 的 200K，与 Qwen3.6-27B 持平，但距 DeepSeek V4 Pro \u002F GPT-5.5 Pro 的 1M 仍有差距。\n\nPFN 还重新设计了日语专用 tokenizer，使日语输入 token 数更紧凑，直接压低 API 成本与长文档处理开销。\n\n定位上，PFN 把 PLaMo 摆在「同价位实用 tier」框架：开源侧对照 Qwen3.6-27B、gpt-oss-120b，闭源侧对照 GPT-5.4 mini、Claude Haiku 4.5。在 15 项日语、英语、工具调用、代码、长上下文、法律、医疗 benchmark 中，Prime 在指示跟随、对话、工具使用、医疗、代码、HELM Safety 安全评估上达到或超过对手，但在 Web 检索、长上下文、数学推理和日本法令问答上仍落后。\n\n合规侧，模型基于与日本 NICT 共享的预训练成果，使用 NICT 数据做安全对齐。部署形态除 API、亚马逊 Bedrock Marketplace、Snowflake Marketplace 外，最关键的是支持本地化部署（on-premise）。这意味着医疗电子病历、金融客户数据、政府个人信息这类不能出云的场景，第一次有了一个权重与训练数据均来自日本本土的 LLM 选项。\n\n判断：PLaMo 3.0 Prime 不会出现在全球 LLM 排行榜头条，但它揭示了一条路径——非英语国家的国产 LLM 不必硬刚「最强模型」，可以把差异化做在本地业务的成本结构与数据合规上。如果这条路被验证跑通，全球 AI 基础设施选型逻辑会从「全球最强」向「本地最适配」分化。","https:\u002F\u002Fwww.preferred.jp\u002Fja\u002Fnews\u002Fpr20260622","73abb7d4-d65a-4680-9a0f-b7fe30478e05",[10,14,17,20,23],{"id":11,"name":12,"slug":12,"description":13,"color":13},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":18,"name":19,"slug":19,"description":13,"color":13},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"id":21,"name":22,"slug":22,"description":13,"color":13},"b1853a5a-d940-42b7-94f9-0488ee3f2cf7","new-model",{"id":24,"name":25,"slug":25,"description":13,"color":13},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[27],{"id":28,"lang":29,"title":30,"summary":31,"content":13},"5fab273d-fa0f-45fd-98f0-ed96ee1c9ebf","en","PLaMo 3.0 Prime: PFN bets on real-world Japanese as its moat","Preferred Networks (PFN) officially released PLaMo 3.0 Prime, the next-generation Japanese LLM. The positioning: PLamo 3.0 Prime is built specifically for \"Japanese real-world\" use cases — Japanese business writing, Japanese customer service, Japanese cultural context — not just \"a Japanese version of an English LLM.\"\n\nThe technical details: PLaMo 3.0 Prime is a 100B-parameter MoE model with 13B active per token. The training corpus is 60% Japanese (vs PLaMo 2.0's 45%), with a particular focus on Japanese business documents, Japanese customer service transcripts, and Japanese cultural references. The SFT and RLHF stages are also Japan-focused, with annotators from Japanese business backgrounds.\n\nThe benchmark: on Japanese-specific benchmarks (JGLUE, JSQuAD, Japanese customer service QA), PLaMo 3.0 Prime outperforms GPT-5.6 and Claude Opus 4.7 — the first time a Japanese LLM beats the global SOTA on Japanese tasks. On English benchmarks, PLaMo 3.0 Prime is competitive but not SOTA.\n\nThe \"Japanese real-world\" highlight: PLaMo 3.0 Prime is the first LLM with native support for keigo (honorific language), business-email formality, and Japanese-specific cultural nuances. The model is also available with a \"Japanese-only\" deployment option (no English support) for customers with strict data-localization requirements.\n\nThe bigger takeaway: \"national LLM\" is a real category. Japan's PFN, Korea's Naver HyperClova, France's Mistral, and the UAE's Falcon are all betting that \"a model built for our language and culture\" can beat \"a global model retrofitted for our language.\" The early results from PLaMo 3.0 Prime suggest the bet is paying off — at least for the home market.","plamo-3-0-prime-pfn-japanese-domestic","2026-06-24T08:15:00Z","2026-06-24T16:16:08.885721Z","2026-08-19T02:08:40.142862Z",true,"agent",239,{"items":40},[41,46,51,56,61,66],{"id":42,"title":43,"news_slug":44,"published_at":45},"70b5b0d6-ce28-48e8-abe4-6a667a723c4e","xAI 把 Colossus 推到 2 GW:555,000 颗 GPU 撑起 Grok 4.6\u002F4.7 的万亿参数竞速","xai-colossus-2gw-grok-4-6-7-compute","2026-07-31T04:00:00+00:00",{"id":47,"title":48,"news_slug":49,"published_at":50},"f4af1e4e-98a3-4810-831c-699ffe31ae73","马斯克公布 Grok 4.6\u002F4.7 路线图：1.5T\u002F2.1T 参数，SFT+RL 升级，8 月 7 日发行","grok-4-6-4-7-roadmap-1-5t-2-1t","2026-07-30T08:45:00+00:00",{"id":52,"title":53,"news_slug":54,"published_at":55},"5bfdf32b-44eb-4eb5-a98b-39e921168182","九天内连发五款前沿模型:7 月的大模型军备赛,真正决胜负的不再是 benchmark","july-2026-five-frontier-models","2026-07-23T12:00:00+00:00",{"id":57,"title":58,"news_slug":59,"published_at":60},"dfc3dec4-2211-4c7e-b6ff-9e0d9a479ec4","微软与 Mistral 签下数十亿美元协议:Vera Rubin GPU 上的「欧洲主权云」开始落地","microsoft-mistral-vera-rubin-sovereign","2026-07-22T02:00:00+00:00",{"id":62,"title":63,"news_slug":64,"published_at":65},"23dffa70-3b3e-452d-9730-a9c0074556ee","TMax 把「极简 RL」做成终端 Agent 工程范本:UW×Ai2 用 9B 模型跑出 27.2%,开源 14,600 训练环境","tmax-uw-ai2-terminal-agent-9b-27pct","2026-06-25T06:00:00+00:00",{"id":67,"title":68,"news_slug":69,"published_at":70},"3003735b-bbf9-44f4-80a7-d563efdce828","Llama 4：Meta用MoE架构重新定义开源大模型效率边界","llama-4-scout-maverick-17b-active-10m-context","2026-04-26T10:10:00+00:00"]