[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news:sw-apple-foundation-models-framework-2026-06-09":3},{"meta":4,"markdown":67},{"type":5,"articleId":6,"slug":7,"title":8,"titleEn":9,"category":10,"summary":11,"publishedAt":12,"image":13,"tags":14,"vocabulary":17,"source":65},"news","news-sw-apple-foundation-models-framework-2026-06-09","sw-apple-foundation-models-framework-2026-06-09","アップル、端末内AIを使うFoundation Models基盤をWWDC26で拡大","Apple Expands Its On-Device Foundation Models Stack for Developers at WWDC26","tech-software","Apple said on June 8 that its third-generation Apple Foundation Models would power a broader Apple Intelligence stack, including developer access through the Foundation Models framework and the new Xcode 27 beta. The announcement matters because it turns Apple's on-device model work into a more practical application platform, with Swift access, built-in tool calling, and privacy-focused execution across on-device and Private Cloud Compute paths. For software engineers, the key question is no longer whether Apple has a model, but how much useful app logic can move onto that stack without adding a separate cloud AI dependency.\n","2026-06-09T00:00:00Z","https:\u002F\u002Fimages.yamiyomi.com\u002Fnews\u002Fsw-apple-foundation-models-framework-2026-06-09.png",[15,16],"ai","us",[18,23,28,32,37,41,45,49,53,57,61],{"word":19,"reading":20,"meaning":21,"level":22},"基盤","きばん","foundation","N2",{"word":24,"reading":25,"meaning":26,"level":27},"端末内","たんまつない","on-device","N1",{"word":29,"reading":30,"meaning":31,"level":27},"推論","すいろん","inference",{"word":33,"reading":34,"meaning":35,"level":36},"開発者","かいはつしゃ","developer","N3",{"word":38,"reading":39,"meaning":40,"level":36},"拡大","かくだい","expansion",{"word":42,"reading":43,"meaning":44,"level":22},"生成","せいせい","generation",{"word":46,"reading":47,"meaning":48,"level":27},"誘導","ゆうどう","guidance",{"word":50,"reading":51,"meaning":52,"level":36},"接続","せつぞく","connection",{"word":54,"reading":55,"meaning":56,"level":27},"疎構造","そこうぞう","sparse architecture",{"word":58,"reading":59,"meaning":60,"level":22},"圧縮","あっしゅく","compression",{"word":62,"reading":63,"meaning":64,"level":36},"検証","けんしょう","verification",{"url":66},"https:\u002F\u002Fmachinelearning.apple.com\u002Fresearch\u002Fintroducing-third-generation-of-apple-foundation-models","\n::heading\n[アップル]{あっぷる:Apple}、[端末内]{たんまつない:on-device:N1}AIを[使う]{つかう:use:N4}Foundation Models[基盤]{きばん:foundation platform:N1}をWWDC26で[拡大]{かくだい:expand:N1}\n\n#en\nApple Expands Its On-Device Foundation Models Platform at WWDC26\n::\n\n::para\n[アップル]{あっぷる:Apple}は[6月8日]{ろくがつようか:June 8:N5}、[第3世代]{だいさんせだい:third generation:N1}のApple Foundation Modelsを[発表]{はっぴょう:announce:N3}し、[同時]{どうじ:at the same time:N4}にFoundation Models frameworkとXcode 27 betaを[通じて]{つうじて:through:N4}、[開発者]{かいはつしゃ:developers:N4}がその[基盤]{きばん:platform:N1}を[使う]{つかう:use:N4}[道]{みち:path:N4}を[広げた]{ひろげた:broadened:N4}。\n\n#en\nOn June 8, Apple introduced its third-generation Apple Foundation Models and at the same time broadened the path for developers to use that stack through the Foundation Models framework and Xcode 27 beta.\n::\n\n::para\n[今回]{こんかい:this time:N3}の[話]{はなし:story:N5}は、[新しい]{あたらしい:new:N4}モデルそのものよりも、[端末内]{たんまつない:on-device:N1}の[大規模言語モデル]{だいきぼげんごもでる:large language model:N1}を[普通]{ふつう:ordinary:N2}のアプリ[機能]{きのう:feature:N3}へどう[組み込む]{くみこむ:embed:N3}かが[見えた]{みえた:became visible:N5}[点]{てん:point:N3}で、ソフトウェア[系]{けい:field:N1}の[読者]{どくしゃ:readers:N4}にとって[重要]{じゅうよう:important:N3}だ。\n\n#en\nMore than the model itself, the important point for software-focused readers is that Apple showed more clearly how an on-device large language model can be embedded into ordinary app features.\n::\n\n::heading\nFoundation Models frameworkの[要点]{ようてん:key points:N3}\n\n#en\nKey Points of the Foundation Models Framework\n::\n\n::para\nAppleの[開発者]{かいはつしゃ:developers:N4}[文書]{ぶんしょ:documentation:N4}によると、このframeworkはApple Intelligenceの[中核]{ちゅうかく:core:N1}にある[端末内]{たんまつない:on-device:N1}[モデル]{もでる:model}へ[アクセス]{あくせす:access}するためのものだ。[記事]{きじ:article:N3}や[日記]{にっき:journal:N3}の[要約]{ようやく:summary:N3}、[質問]{しつもん:question:N4}への[応答]{おうとう:response:N1}、アプリ[内]{ない:inside:N3}の[作業]{さぎょう:tasks:N4}の[実行]{じっこう:execution:N3}などを、Swiftから[近い]{ちかい:direct:N4}かたちで[呼び出せる]{よびだせる:can call:N3}。\n\n#en\nAccording to Apple's developer documentation, the framework gives access to the on-device model at the core of Apple Intelligence. It lets apps call that model from Swift for tasks such as summarization, answering questions, and carrying out in-app work.\n::\n\n::para\n[加えて]{くわえて:in addition:N3}、[誘導生成]{ゆうどうせいせい:guided generation:N1}やtool callingが[組み込まれている]{くみこまれている:built in:N3}ため、[開発者]{かいはつしゃ:developers:N4}は[単なる]{たんなる:mere:N3}[文章生成]{ぶんしょうせいせい:text generation:N2}だけでなく、アプリの[内部]{ないぶ:inside:N3}データや[操作]{そうさ:actions:N1}を[接続]{せつぞく:connect:N2}した[設計]{せっけい:design:N2}を[試しやすい]{ためしやすい:easier to try:N4}と[見られる]{みられる:is seen:N5}。\n\n#en\nBecause guided generation and tool calling are built in, developers appear able to try designs that connect the model to internal app data and actions rather than using it only for plain text generation.\n::\n\n::callout\n[昨年]{さくねん:last year:N3}からAppleは、[端末内]{たんまつない:on-device:N1}[推論]{すいろん:inference:N1}は[オフライン]{おふらいん:offline}でも[使える]{つかえる:usable:N4}こと、[推論]{すいろん:inference:N1}[費用]{ひよう:cost:N3}を[別の]{べつの:separate:N4}クラウドAIサービスのように[払わない]{はらわない:not pay:N3}でよいこと、そして[個人]{こじん:personal:N2}[情報]{じょうほう:information:N3}を[外部]{がいぶ:outside:N3}へ[出しにくい]{だしにくい:harder to send out:N5}ことを[強み]{つよみ:strengths:N4}として[示してきた]{しめしてきた:has presented:N3}。[今回]{こんかい:this time:N3}はその[路線]{ろせん:line:N2}をさらに[強めた]{つよめた:strengthened:N4}かたちだ。\n\n#en\nSince last year, Apple has presented offline use, no separate inference bill, and lower pressure to send personal data outward as strengths of its on-device approach. This announcement strengthens that line.\n::\n\n::heading\n[モデル]{もでる:model}[側]{がわ:side:N3}で[何]{なに:what:N5}が[変わった]{かわった:changed:N3}か\n\n#en\nWhat Changed on the Model Side\n::\n\n::para\nAppleの[研究]{けんきゅう:research:N4}[説明]{せつめい:explanation:N3}では、[第3世代]{だいさんせだい:third generation:N1}の[構成]{こうせい:architecture:N3}は5つのモデル[族]{ぞく:family:N4}からなり、2つが[端末内]{たんまつない:on-device:N1}、3つがPrivate Cloud Compute[向け]{むけ:for:N3}だという。[端末内]{たんまつない:on-device:N1}のAFM 3 Coreは30[億]{おく:billion:N2}パラメーターのdense modelで、[前世代]{ぜんせだい:previous generation:N4}より[品質]{ひんしつ:quality:N4}が[上がった]{あがった:improved:N5}とされた。\n\n#en\nApple's research post says the third-generation stack is a family of five models, with two on-device and three for Private Cloud Compute. The on-device AFM 3 Core is a 3-billion-parameter dense model that Apple says improves on the previous generation.\n::\n\n::para\nもう1つのAFM 3 Core Advancedは、20[億]{おく:billion:N2}ではなく200[億]{おく:billion:N2}パラメーター[級]{きゅう:class:N1}の[疎構造]{そこうぞう:sparse architecture:N1}[モデル]{もでる:model}で、[要求]{ようきゅう:request:N3}ごとに10[億]{おく:billion:N2}から40[億]{おく:billion:N2}パラメーターを[動かす]{うごかす:activate:N4} [設計]{せっけい:design:N2}だという。[音声]{おんせい:voice:N3}や[高精度]{こうせいど:high accuracy:N3}の[書き取り]{かきとり:dictation:N3}のような[用途]{ようと:use cases:N3}を[支える]{ささえる:support:N3}と[説明]{せつめい:explain:N3}された。\n\n#en\nThe other on-device model, AFM 3 Core Advanced, is described as a 20-billion-parameter sparse model that activates 1 billion to 4 billion parameters depending on the request. Apple said it supports uses such as expressive voice features and higher-accuracy dictation.\n::\n\n::para\n[クラウド]{くらうど:cloud}[側]{がわ:side:N3}では、AFM 3 Cloud、[画像]{がぞう:images:N2}[向け]{むけ:for:N3}のADM 3 Cloud、そしてAFM 3 Cloud Proが[並ぶ]{ならぶ:line up:N2}。AppleはAFM 3 Cloud Proについて、Google Cloud[上]{じょう:on top of:N5}のNVIDIA GPUでもPrivate Cloud Computeと[同じ]{おなじ:same:N4}[保護]{ほご:protections:N1}を[保つ]{たもつ:maintain:N1}よう[拡張]{かくちょう:extend:N1}したと[述べた]{のべた:stated:N2}。\n\n#en\nOn the cloud side, Apple lists AFM 3 Cloud, the image-focused ADM 3 Cloud, and AFM 3 Cloud Pro. Apple said AFM 3 Cloud Pro extends Private Cloud Compute protections onto NVIDIA GPUs in Google Cloud.\n::\n\n::heading\n[開発者]{かいはつしゃ:developers:N4}が[今]{いま:now:N5}[見る]{みる:watch:N5}べきこと\n\n#en\nWhat Developers Should Watch Now\n::\n\n::para\n[同じ]{おなじ:same:N4}[6月8日]{ろくがつようか:June 8:N5}に[公開]{こうかい:released:N4}されたXcode 27 betaは、Swift 6.4とiOS 27などのSDKを[含む]{ふくむ:include:N2}。[最新]{さいしん:latest:N3}の[開発環境]{かいはつかんきょう:development environment:N1}とFoundation Models frameworkが[並んだ]{ならんだ:arrived together:N2}ことで、Appleは[端末内]{たんまつない:on-device:N1}AIを[研究]{けんきゅう:research:N4}の[題材]{だいざい:subject:N2}ではなく、[実装]{じっそう:implementation:N2}の[対象]{たいしょう:target:N2}へ[押し出した]{おしだした:pushed:N3}と[言えそう]{いえそう:it seems fair to say:N4}だ。\n\n#en\nXcode 27 beta, released the same day, includes Swift 6.4 and the new platform SDKs. Arriving alongside the Foundation Models framework, it pushes on-device AI from a research topic toward something developers can actually implement.\n::\n\n::para\nただし、Appleの[文書]{ぶんしょ:documentation:N4}では、[利用者]{りようしゃ:users:N3}がApple Intelligenceを[有効化]{ゆうこうか:enable:N2}したあと、[モデル]{もでる:model}の[ダウンロード]{だうんろーど:download}に[時間]{じかん:time:N5}がかかることがあるとも[書かれている]{かかれている:is written:N5}。[開発者]{かいはつしゃ:developers:N4}は、[応答]{おうとう:responses:N1}の[速さ]{はやさ:speed:N3}だけでなく、[初回]{しょかい:first-time:N3}の[準備]{じゅんび:setup:N2}、[対応端末]{たいおうたんまつ:supported devices:N1}、[検証]{けんしょう:verification:N1}の[流れ]{ながれ:workflow:N3}まで[含めて]{ふくめて:including:N2} [設計]{せっけい:design:N2}する[必要]{ひつよう:need:N3}がある。\n\n#en\nApple's documentation also notes that after a user enables Apple Intelligence, it can take time for the model to download. Developers therefore need to design not only for response speed, but also for first-run setup, device support, and testing workflow.\n::\n"]