{"@context":"https://schema.org","@type":"NewsArticle","generatedAt":"2026-08-11T09:21:12.743Z","headline":"苹果 iPhone 上最快本地 AI 模型：Maple-Preview-20B-A1B 登场，可达 127 tokens/s","description":"DeepGrove 发布轻量 AI 模型 Maple-Preview，在 iPhone 上运行速度达 127 tokens/s，约为 Bonsai 27B 的 13 倍。该模型总参数量 202 亿，采用 MoE 架构，激活参数 14.9 亿，并提出\"三值权重\"方案降低内存占用。目前处于预览阶段，处理 131，000 tokens 时内存占用仅 7.69GB。","url":"https://www.aioga.com/news/cmsh6fjwm002hronk4z70xx1v/","mainEntityOfPage":"https://www.aioga.com/news/cmsh6fjwm002hronk4z70xx1v/","datePublished":"2026-08-06T06:21:57.000Z","dateModified":"2026-08-06T06:21:57.000Z","inLanguage":"zh-CN","publisher":{"@type":"NewsMediaOrganization","name":"Aioga","url":"https://www.aioga.com"},"citation":["https://www.ithome.com/0/986/493.htm","https://aihot.virxact.com/items/cmsh6fjwm002hronk4z70xx1v"],"canonicalUrl":"https://www.aioga.com/news/cmsh6fjwm002hronk4z70xx1v/","directAnswer":{"@type":"Answer","text":"DeepGrove 发布预览阶段轻量模型 Maple-Preview。按材料所述，该模型在未明确具体机型的 iPhone 上运行速度可达 127 tokens/s；总参数量为 202 亿，采用 MoE 架构，激活参数量为 14.9 亿。","url":"https://www.aioga.com/news/cmsh6fjwm002hronk4z70xx1v/","dateCreated":"2026-08-06T06:21:57.000Z","author":{"@type":"Organization","@id":"https://www.aioga.com/authors/aioga-editorial/#editorial-team","name":"Aioga Editorial Team","url":"https://www.aioga.com/authors/aioga-editorial/"}},"evidence":[{"@type":"CreativeWork","name":"IT之家 source article","url":"https://www.ithome.com/0/986/493.htm","datePublished":"2026-08-06T06:21:57.000Z","provider":{"@type":"Organization","name":"IT之家","url":"https://www.ithome.com/0/986/493.htm"}},{"@type":"CreativeWork","name":"AIHot archive record","url":"https://aihot.virxact.com/items/cmsh6fjwm002hronk4z70xx1v","datePublished":"2026-08-06T06:21:57.000Z","provider":{"@type":"Organization","name":"AIHot","url":"https://aihot.virxact.com/items/cmsh6fjwm002hronk4z70xx1v"}}],"aggregationSource":"IT之家（RSS）","originalPublisher":{"name":"IT之家","url":"https://www.ithome.com/0/986/493.htm"},"geoDeepAnswer":null,"article":{"id":"cmsh6fjwm002hronk4z70xx1v","slug":"cmsh6fjwm002hronk4z70xx1v","url":"https://www.aioga.com/news/cmsh6fjwm002hronk4z70xx1v/","title":"苹果 iPhone 上最快本地 AI 模型：Maple-Preview-20B-A1B 登场，可达 127 tokens/s","title_en":"","summary":"DeepGrove 发布轻量 AI 模型 Maple-Preview，在 iPhone 上运行速度达 127 tokens/s，约为 Bonsai 27B 的 13 倍。该模型总参数量 202 亿，采用 MoE 架构，激活参数 14.9 亿，并提出\"三值权重\"方案降低内存占用。目前处于预览阶段，处理 131，000 tokens 时内存占用仅 7.69GB。","source":"IT之家（RSS）","sourceUrl":"https://www.ithome.com/0/986/493.htm","aiHotUrl":"https://aihot.virxact.com/items/cmsh6fjwm002hronk4z70xx1v","publishedAt":"2026-08-06T06:21:57.000Z","category":"模型更新","score":52,"selected":false,"articleBody":["IT之家：https://www.ithome.com/ 8 月 6 日消息，美国独立 AI 研究实验室 DeepGrove 昨日（8 月 5 日）发布博文，发布名为 Maple-Preview 的轻量 AI 模型， 在 iPhone：https://iphone.ithome.com/（原文并未明确具体机型）上运行可以达到 127 tokens/s，是 Bonsai 27B (9.6 tokens/s) 的约 13 倍：https://www.ithome.com/0/976/772.htm。","在模型规模方面，Maple-Preview 的总参数量为 202 亿个，采用混合专家（MoE）架构，激活参数量为 14.9 亿个。","DeepGrove 表示，目前 AI 行业针对本地端侧主要依赖量化技术，但这种方式存在根本性问题，严重影响模型的性能和效率。","DeepGrove 公司提出“三值权重”（ternary-weight）方案，将权重约束并量化为三值（{-1, 0, 1})），大幅降低了内存占用与计算带宽需求，该模型包含 24 层、256 个专家（8 个活跃专家），并采用 3:1 SWA-512:GA 混合注意力机制。","在能力展示上，Maple-Preview 在 MacBook Pro（M5 Pro）上完成 IMO 2024 P1，结果为 7/7，运行速度 281.5 tokens/s。","在长上下文场景下，Maple-Preview 也强调内存控制能力。相关图表把 context length（上下文长度）与维持上下文所需的额外内存占用放在同一坐标系中。结果显示，即便处理 131,000 tokens，Maple-Preview 的内存占用仍只有 7.69GB。","DeepGrove 表示，Maple-Preview 目前仍处于 preview stage（预览阶段），后续还会继续改进。公司同时希望构建一套系统，让 AI 模型能根据对话内容自动调整，并针对单个用户做优化。"],"articleImages":[{"sourceUrl":"https://img.ithome.com/newsuploadfiles/2026/8/c790d444-0c68-4900-b186-0eb65e225721.png","alt":"On-device decode speed versus benchmark performance for Maple-Preview with its dense and Flash heads and comparison models.","afterParagraph":3,"url":"/media/articles/cmsh6fjwm002hronk4z70xx1v/10acfd2c9cd00f78.png"},{"sourceUrl":"https://img.ithome.com/newsuploadfiles/2026/8/0775849a-32f7-4c98-92bd-9b30d2d2840d.jpg@s_2,w_820,h_352","alt":"","afterParagraph":4,"url":"/media/articles/cmsh6fjwm002hronk4z70xx1v/d09edd218cffbcde.jpg"},{"sourceUrl":"https://img.ithome.com/newsuploadfiles/2026/8/72768272-eaeb-4ad9-9efc-94df363a8edf.png","alt":"Projected resident model-weight and context-state memory at 131,072 tokens for Maple and comparison models.","afterParagraph":5,"url":"/media/articles/cmsh6fjwm002hronk4z70xx1v/da7b72318b7f0f1f.png"},{"sourceUrl":"https://img.ithome.com/newsuploadfiles/2026/8/340bee7d-7cff-44c1-9980-b2749d4925a6.png","alt":"Incremental context-memory growth above a 640-token baseline through 131,072 tokens for Maple and comparison architectures.","afterParagraph":5,"url":"/media/articles/cmsh6fjwm002hronk4z70xx1v/2de2471f09cb859b.png"},{"sourceUrl":"https://img.ithome.com/newsuploadfiles/2026/8/c8ac220d-6cee-4d29-b0f5-98aa44ceede3.png","alt":"Grouped benchmark scores for Maple-Preview and four comparison models across LCBv6, AIME 2026, HMMT 2026, GPQA-D, and the average, using a consistent model order with Maple first.","afterParagraph":5,"url":"/media/articles/cmsh6fjwm002hronk4z70xx1v/18c2dc64280248c1.png"},{"sourceUrl":"https://img.ithome.com/newsuploadfiles/2026/8/abbdce8c-5a70-478d-a7c6-b391be220990.png","alt":"","afterParagraph":5,"url":"/media/articles/cmsh6fjwm002hronk4z70xx1v/194871c63da42b6a.png"}],"mediaStatus":"ok","articleBodyZh":["苹果 iPhone 上最快本地 AI 模型：Maple-Preview-20B-A1B 登场，可达 127 tokens/s 这条更新来自 ithome.com，发布时间为 2026-08-06，Aioga 保留原文入口以便核验。","摘要：DeepGrove 发布轻量 AI 模型 Maple-Preview，在 iPhone 上运行速度达 127 tokens/s，约为 Bonsai 27B 的 13 倍。该模型总参数量 202 亿，采用 MoE 架构，激活参数 14.9 亿，并提出\"三值权重\"方案降低内存占用。目前处于预览阶段，处理 131，000 tokens 时内存占用仅 7.69GB。","背景：材料称，DeepGrove 认为端侧 AI 目前主要依赖量化技术，并指出这会影响模型性能和效率。Maple-Preview 使用三值权重方案，将权重约束并量化为 {-1,0,1}，以降低内存占用和计算带宽需求。","Aioga 观察：Aioga 判断，Maple-Preview 的信息重点不只在单项速度，还在于以较低激活参数量和三值权重方案服务本地运行。其 127 tokens/s 的表述应结合未披露具体 iPhone 机型、仍处预览阶段等条件谨慎看待。","影响与后续：值得关注的是，材料给出的长上下文数据称，处理 131,000 tokens 时内存占用为 7.69GB。若后续测试能够覆盖更多设备与任务，本地模型在速度、内存控制和上下文长度之间的取舍或将更易比较。 下一步可关注 DeepGrove 对预览版本的后续改进，以及其是否披露具体 iPhone 测试机型、测试条件和更多实际任务表现。同时，可留意其关于模型随对话自动调整并面向单个用户优化的系统是否形成可验证成果。"],"translationStatus":"","bodyOrigin":"source-page","editorial":{"summary":"DeepGrove 发布预览阶段轻量模型 Maple-Preview。按材料所述，该模型在未明确具体机型的 iPhone 上运行速度可达 127 tokens/s；总参数量为 202 亿，采用 MoE 架构，激活参数量为 14.9 亿。","background":"材料称，DeepGrove 认为端侧 AI 目前主要依赖量化技术，并指出这会影响模型性能和效率。Maple-Preview 使用三值权重方案，将权重约束并量化为 {-1,0,1}，以降低内存占用和计算带宽需求。","viewpoint":"Aioga 判断，Maple-Preview 的信息重点不只在单项速度，还在于以较低激活参数量和三值权重方案服务本地运行。其 127 tokens/s 的表述应结合未披露具体 iPhone 机型、仍处预览阶段等条件谨慎看待。","implications":"值得关注的是，材料给出的长上下文数据称，处理 131,000 tokens 时内存占用为 7.69GB。若后续测试能够覆盖更多设备与任务，本地模型在速度、内存控制和上下文长度之间的取舍或将更易比较。","nextStep":"下一步可关注 DeepGrove 对预览版本的后续改进，以及其是否披露具体 iPhone 测试机型、测试条件和更多实际任务表现。同时，可留意其关于模型随对话自动调整并面向单个用户优化的系统是否形成可验证成果。","evidenceRefs":["title","summary","articleBody","source"],"status":"published","aiGenerated":true,"autoApproved":true,"generatedBy":"aioga-editorial:gpt-5.6-sol","reviewedBy":"aioga-editorial-review:gpt-5.6-sol","generatedAt":"2026-08-10T08:03:48.271Z","sourceHash":"0b899be8d636f01d","review":{"approved":true,"groundedness":96,"clarity":91,"duplicationRisk":18,"blockingIssues":[],"notes":["候选内容准确保留了“未明确具体 iPhone 机型”和“预览阶段”等限制条件，避免将 127 tokens/s 表述为普遍适用于所有 iPhone 的结果。","“若后续测试能够覆盖更多设备与任务”等内容属于明确的条件性判断，没有冒充已发生事实。","“Aioga 判断”中的观点与事实陈述区分清楚；如需提升可读性，可补充 Aioga 的身份或改为更中性的“综合材料判断”，但这不构成阻断问题。"]},"validation":{"passed":true,"mode":"ai-auto","revisions":0,"checks":["schema","length","source-attribution","low-source-overlap","no-html","independent-ai-review"]}},"tags":["模型更新","IT之家（RSS）"],"translations":{"zh-CN":{"title":"苹果 iPhone 上最快本地 AI 模型：Maple-Preview-20B-A1B 登场，可达 127 tokens/s","summary":"DeepGrove 发布轻量 AI 模型 Maple-Preview，在 iPhone 上运行速度达 127 tokens/s，约为 Bonsai 27B 的 13 倍。该模型总参数量 202 亿，采用 MoE 架构，激活参数 14.9 亿，并提出\"三值权重\"方案降低内存占用。目前处于预览阶段，处理 131，000 tokens 时内存占用仅 7.69GB。","category":"模型更新","source":"IT之家","aggregationSource":"IT之家（RSS）","pageTitle":"苹果 iPhone 上最快本地 AI 模型：Maple-Preview-20B-A1B 登场，可达 127 tokens/s - Aioga AI资讯","description":"DeepGrove 发布轻量 AI 模型 Maple-Preview，在 iPhone 上运行速度达 127 tokens/s，约为 Bonsai 27B 的 13 倍。该模型总参数量 202 亿，采用 MoE 架构，激活参数 14.9 亿，并提出\"三值权重\"方案降低内存占用。目前处于预览阶段，处理 131，000 tokens 时内存占用仅 7.69GB。","url":"https://www.aioga.com/news/cmsh6fjwm002hronk4z70xx1v/","articleBody":["苹果 iPhone 上最快本地 AI 模型：Maple-Preview-20B-A1B 登场，可达 127 tokens/s 这条更新来自 ithome.com，发布时间为 2026-08-06，Aioga 保留原文入口以便核验。","摘要：DeepGrove 发布轻量 AI 模型 Maple-Preview，在 iPhone 上运行速度达 127 tokens/s，约为 Bonsai 27B 的 13 倍。该模型总参数量 202 亿，采用 MoE 架构，激活参数 14.9 亿，并提出\"三值权重\"方案降低内存占用。目前处于预览阶段，处理 131，000 tokens 时内存占用仅 7.69GB。","背景：材料称，DeepGrove 认为端侧 AI 目前主要依赖量化技术，并指出这会影响模型性能和效率。Maple-Preview 使用三值权重方案，将权重约束并量化为 {-1,0,1}，以降低内存占用和计算带宽需求。","Aioga 观察：Aioga 判断，Maple-Preview 的信息重点不只在单项速度，还在于以较低激活参数量和三值权重方案服务本地运行。其 127 tokens/s 的表述应结合未披露具体 iPhone 机型、仍处预览阶段等条件谨慎看待。","影响与后续：值得关注的是，材料给出的长上下文数据称，处理 131,000 tokens 时内存占用为 7.69GB。若后续测试能够覆盖更多设备与任务，本地模型在速度、内存控制和上下文长度之间的取舍或将更易比较。 下一步可关注 DeepGrove 对预览版本的后续改进，以及其是否披露具体 iPhone 测试机型、测试条件和更多实际任务表现。同时，可留意其关于模型随对话自动调整并面向单个用户优化的系统是否形成可验证成果。"]},"en":{"title":"The fastest local AI model on Apple iPhone: Maple-Preview-20B-A1B launched, reaching 127 tokens/s","summary":"DeepGrove released the lightweight AI model Maple-Preview, running at a speed of 127 tokens/s on the iPhone, about 13 times that of Bonsai 27B. The model has a total of 20.2 billion parameters, uses an MoE architecture, has 1.49 billion active parameters, and proposes a 'ternary weight' scheme to reduce memory usage. Currently in the preview stage, it only consumes 7.69GB of memory when processing 131,000 tokens.","category":"Models","source":"IT之家（RSS）","aggregationSource":"IT之家（RSS）","pageTitle":"The fastest local AI model on Apple iPhone: Maple-Preview-20B-A1B launched, reaching 127 tokens/s - Aioga AI News","description":"DeepGrove released the lightweight AI model Maple-Preview, running at a speed of 127 tokens/s on the iPhone, about 13 times that of Bonsai 27B. The model has a total of 20.2 billio...","url":"https://www.aioga.com/en/news/cmsh6fjwm002hronk4z70xx1v/","contentTranslated":true,"sourceHash":"b1d27ba2a743f55e","translatedAt":"2026-08-06T11:02:54.882Z"},"ja":{"title":"Apple iPhone 上で最速のローカル AI モデル：Maple-Preview-20B-A1B 登場、127 tokens/s に達する","summary":"DeepGrove は軽量 AI モデル Maple-Preview を発表し、iPhone 上での動作速度は 127 tokens/s に達し、Bonsai 27B の約 13 倍です。このモデルの総パラメータ数は 202 億で、MoE アーキテクチャを採用し、アクティブパラメータは 14.9 億で、「三値重み」方式を提案してメモリ使用量を削減しています。現在はプレビュー段階であり、131,000 tokens を処理する際のメモリ使用量はわずか 7.69GB です。","category":"モデル更新","source":"IT之家（RSS）","aggregationSource":"IT之家（RSS）","pageTitle":"Apple iPhone 上で最速のローカル AI モデル：Maple-Preview-20B-A1B 登場、127 tokens/s に達する - Aioga AIニュース","description":"DeepGrove は軽量 AI モデル Maple-Preview を発表し、iPhone 上での動作速度は 127 tokens/s に達し、Bonsai 27B の約 13 倍です。このモデルの総パラメータ数は 202 億で、MoE アーキテクチャを採用し、アクティブパラメータは 14.9 億で、「三値重み」方式を提案してメモリ使用量を削減しています。...","url":"https://www.aioga.com/ja/news/cmsh6fjwm002hronk4z70xx1v/","contentTranslated":true,"sourceHash":"b1d27ba2a743f55e","translatedAt":"2026-08-06T11:03:08.794Z"},"ko":{"title":"애플 iPhone에서 가장 빠른 로컬 AI 모델: Maple-Preview-20B-A1B 등장, 최대 127 tokens/s","summary":"DeepGrove는 경량 AI 모델 Maple-Preview를 발표했으며, iPhone에서 실행 속도가 127 tokens/s에 달해 Bonsai 27B보다 약 13배 빠릅니다. 이 모델의 총 매개변수 수는 202억이며, MoE 아키텍처를 사용하고 활성화 매개변수는 14.9억입니다. 또한 메모리 사용을 줄이기 위해 '삼진 가중치' 방안을 제안했습니다. 현재 미리보기 단계에 있으며, 131,000 tokens를 처리할 때 메모리 사용량은 단 7.69GB에 불과합니다.","category":"모델 업데이트","source":"IT之家（RSS）","aggregationSource":"IT之家（RSS）","pageTitle":"애플 iPhone에서 가장 빠른 로컬 AI 모델: Maple-Preview-20B-A1B 등장, 최대 127 tokens/s - Aioga AI 뉴스","description":"DeepGrove는 경량 AI 모델 Maple-Preview를 발표했으며, iPhone에서 실행 속도가 127 tokens/s에 달해 Bonsai 27B보다 약 13배 빠릅니다. 이 모델의 총 매개변수 수는 202억이며, MoE 아키텍처를 사용하고 활성화 매개변수는 14.9억입니다. 또한 메모리 사용을 줄이기 위해 '삼진...","url":"https://www.aioga.com/ko/news/cmsh6fjwm002hronk4z70xx1v/","contentTranslated":true,"sourceHash":"b1d27ba2a743f55e","translatedAt":"2026-08-06T11:04:01.993Z"},"es":{"title":"El modelo de IA local más rápido en el iPhone de Apple: Maple-Preview-20B-A1B llega, puede alcanzar 127 tokens/s","summary":"DeepGrove lanzó el modelo de IA ligero Maple-Preview, que funciona a una velocidad de 127 tokens/s en iPhone, aproximadamente 13 veces más rápido que Bonsai 27B. Este modelo tiene un total de 20.2 mil millones de parámetros, utiliza arquitectura MoE, con 1.49 mil millones de parámetros activados, y propone un esquema de \"pesos ternarios\" para reducir el uso de memoria. Actualmente se encuentra en fase de vista previa, y al procesar 131,000 tokens, el uso de memoria es de solo 7.69GB.","category":"Modelos","source":"IT之家（RSS）","aggregationSource":"IT之家（RSS）","pageTitle":"El modelo de IA local más rápido en el iPhone de Apple: Maple-Preview-20B-A1B llega, puede alcanzar 127 tokens/s - Aioga Noticias de IA","description":"DeepGrove lanzó el modelo de IA ligero Maple-Preview, que funciona a una velocidad de 127 tokens/s en iPhone, aproximadamente 13 veces más rápido que Bonsai 27B. Este modelo tiene...","url":"https://www.aioga.com/es/news/cmsh6fjwm002hronk4z70xx1v/","contentTranslated":true,"sourceHash":"b1d27ba2a743f55e","translatedAt":"2026-08-06T11:03:59.021Z"},"fr":{"title":"Le modèle d'IA local le plus rapide sur iPhone : Maple-Preview-20B-A1B arrive, pouvant atteindre 127 tokens/s","summary":"DeepGrove a publié le modèle AI léger Maple-Preview, qui fonctionne sur iPhone à une vitesse de 127 tokens/s, environ 13 fois celle de Bonsai 27B. Ce modèle compte au total 20,2 milliards de paramètres, utilise une architecture MoE avec 1,49 milliard de paramètres activés, et propose une solution de \"poids à trois valeurs\" pour réduire l'utilisation de la mémoire. Il est actuellement en phase de prévisualisation, avec une utilisation de mémoire de seulement 7,69 Go pour traiter 131 000 tokens.","category":"Modèles","source":"IT之家（RSS）","aggregationSource":"IT之家（RSS）","pageTitle":"Le modèle d'IA local le plus rapide sur iPhone : Maple-Preview-20B-A1B arrive, pouvant atteindre 127 tokens/s - Aioga Actualités IA","description":"DeepGrove a publié le modèle AI léger Maple-Preview, qui fonctionne sur iPhone à une vitesse de 127 tokens/s, environ 13 fois celle de Bonsai 27B. Ce modèle compte au total 20,2 mi...","url":"https://www.aioga.com/fr/news/cmsh6fjwm002hronk4z70xx1v/","contentTranslated":true,"sourceHash":"b1d27ba2a743f55e","translatedAt":"2026-08-06T11:04:52.654Z"},"de":{"title":"Das schnellste lokale KI-Modell auf dem Apple iPhone: Maple-Preview-20B-A1B vorgestellt, erreicht bis zu 127 Tokens/s","summary":"DeepGrove hat das leichte KI-Modell Maple-Preview veröffentlicht, das auf dem iPhone mit einer Geschwindigkeit von 127 Tokens/s läuft, etwa 13-mal schneller als Bonsai 27B. Das Modell hat insgesamt 20,2 Milliarden Parameter, verwendet eine MoE-Architektur, 1,49 Milliarden aktivierte Parameter und schlägt das \"Dreiwert-Gewichts\"-Schema vor, um den Speicherbedarf zu reduzieren. Es befindet sich derzeit in der Vorschauphase und benötigt bei der Verarbeitung von 131.000 Tokens nur 7,69 GB Speicher.","category":"模型更新","source":"IT之家（RSS）","aggregationSource":"IT之家（RSS）","pageTitle":"Das schnellste lokale KI-Modell auf dem Apple iPhone: Maple-Preview-20B-A1B vorgestellt, erreicht bis zu 127 Tokens/s - Aioga KI-News","description":"DeepGrove hat das leichte KI-Modell Maple-Preview veröffentlicht, das auf dem iPhone mit einer Geschwindigkeit von 127 Tokens/s läuft, etwa 13-mal schneller als Bonsai 27B. Das Mod...","url":"https://www.aioga.com/de/news/cmsh6fjwm002hronk4z70xx1v/","contentTranslated":true,"sourceHash":"b1d27ba2a743f55e","translatedAt":"2026-08-06T11:04:52.047Z"},"pt-BR":{"title":"O modelo de IA local mais rápido no iPhone da Apple: Maple-Preview-20B-A1B estreando, podendo atingir 127 tokens/s","summary":"DeepGrove lançou o modelo de IA leve Maple-Preview, que atinge uma velocidade de 127 tokens/s no iPhone, cerca de 13 vezes a velocidade do Bonsai 27B. O modelo possui um total de 20,2 bilhões de parâmetros, utiliza arquitetura MoE, com 1,49 bilhão de parâmetros ativados, e propõe o esquema de \"peso ternário\" para reduzir o uso de memória. Atualmente está em fase de pré-visualização, ocupando apenas 7,69 GB de memória ao processar 131.000 tokens.","category":"模型更新","source":"IT之家（RSS）","aggregationSource":"IT之家（RSS）","pageTitle":"O modelo de IA local mais rápido no iPhone da Apple: Maple-Preview-20B-A1B estreando, podendo atingir 127 tokens/s - Aioga Notícias de IA","description":"DeepGrove lançou o modelo de IA leve Maple-Preview, que atinge uma velocidade de 127 tokens/s no iPhone, cerca de 13 vezes a velocidade do Bonsai 27B. O modelo possui um total de 2...","url":"https://www.aioga.com/pt-BR/news/cmsh6fjwm002hronk4z70xx1v/","contentTranslated":true,"sourceHash":"b1d27ba2a743f55e","translatedAt":"2026-08-06T11:05:37.283Z"},"ru":{"title":"Самая быстрая локальная AI-модель на iPhone: Maple-Preview-20B-A1B дебютирует с производительностью до 127 токенов/с","summary":"DeepGrove выпустила легкую AI модель Maple-Preview, которая работает на iPhone со скоростью 127 токенов/с, что примерно в 13 раз быстрее, чем у Bonsai 27B. Общий объем параметров модели составляет 20,2 миллиарда, используется архитектура MoE, активные параметры — 1,49 миллиарда, и предложена схема «тройных весов» для снижения использования памяти. В настоящее время модель находится на этапе превью, и при обработке 131 000 токенов память используется всего на 7,69 ГБ.","category":"模型更新","source":"IT之家（RSS）","aggregationSource":"IT之家（RSS）","pageTitle":"Самая быстрая локальная AI-модель на iPhone: Maple-Preview-20B-A1B дебютирует с производительностью до 127 токенов/с - Aioga Новости ИИ","description":"DeepGrove выпустила легкую AI модель Maple-Preview, которая работает на iPhone со скоростью 127 токенов/с, что примерно в 13 раз быстрее, чем у Bonsai 27B. Общий объем параметров м...","url":"https://www.aioga.com/ru/news/cmsh6fjwm002hronk4z70xx1v/","contentTranslated":true,"sourceHash":"b1d27ba2a743f55e","translatedAt":"2026-08-06T11:05:43.673Z"},"ar":{"title":"أسرع نموذج ذكاء اصطناعي محلي على iPhone من آبل: إطلاق Maple-Preview-20B-A1B، يمكن أن يصل إلى 127 رمزًا/ثانية","summary":"أصدرت DeepGrove نموذج AI خفيف الوزن Maple-Preview، بسرعة تشغيل على iPhone تصل إلى 127 رمزًا/ثانية، أي بما يقارب 13 ضعفًا لموديل Bonsai 27B. يبلغ إجمالي عدد المعلمات في هذا النموذج 20.2 مليار، ويعتمد على بنية MoE، مع 1.49 مليار معلمة مفعلة، واقترح \"خطة الوزن ثلاثي القيم\" لتقليل استهلاك الذاكرة. حالياً هو في المرحلة التجريبية، وعند معالجة 131,000 رمز، يبلغ استهلاك الذاكرة فقط 7.69 جيجابايت.","category":"模型更新","source":"IT之家（RSS）","aggregationSource":"IT之家（RSS）","pageTitle":"أسرع نموذج ذكاء اصطناعي محلي على iPhone من آبل: إطلاق Maple-Preview-20B-A1B، يمكن أن يصل إلى 127 رمزًا/ثانية - Aioga أخبار الذكاء الاصطناعي","description":"أصدرت DeepGrove نموذج AI خفيف الوزن Maple-Preview، بسرعة تشغيل على iPhone تصل إلى 127 رمزًا/ثانية، أي بما يقارب 13 ضعفًا لموديل Bonsai 27B. يبلغ إجمالي عدد المعلمات في هذا النموذج...","url":"https://www.aioga.com/ar/news/cmsh6fjwm002hronk4z70xx1v/","contentTranslated":true,"sourceHash":"b1d27ba2a743f55e","translatedAt":"2026-08-06T11:06:33.974Z"},"hi":{"title":"एप्पल iPhone पर सबसे तेज़ लोकल AI मॉडल: Maple-Preview-20B-A1B पेश, 127 tokens/s तक पहुँच सकता है","summary":"DeepGrove ने हल्का AI मॉडल Maple-Preview जारी किया, जो iPhone पर 127 tokens/s की गति से चलता है, जो Bonsai 27B से लगभग 13 गुना तेज है। इस मॉडल में कुल 20.2 बिलियन पैरामीटर हैं, यह MoE आर्किटेक्चर का उपयोग करता है, सक्रिय पैरामीटर 1.49 बिलियन हैं, और मेमोरी उपयोग कम करने के लिए 'तीन-मूल्य वजन' योजना पेश की गई है। वर्तमान में यह प्रीव्यू चरण में है, और 131,000 tokens को प्रोसेस करते समय मेमोरी उपयोग केवल 7.69GB है।","category":"模型更新","source":"IT之家（RSS）","aggregationSource":"IT之家（RSS）","pageTitle":"एप्पल iPhone पर सबसे तेज़ लोकल AI मॉडल: Maple-Preview-20B-A1B पेश, 127 tokens/s तक पहुँच सकता है - Aioga AI समाचार","description":"DeepGrove ने हल्का AI मॉडल Maple-Preview जारी किया, जो iPhone पर 127 tokens/s की गति से चलता है, जो Bonsai 27B से लगभग 13 गुना तेज है। इस मॉडल में कुल 20.2 बिलियन पैरामीटर हैं, यह...","url":"https://www.aioga.com/hi/news/cmsh6fjwm002hronk4z70xx1v/","contentTranslated":true,"sourceHash":"b1d27ba2a743f55e","translatedAt":"2026-08-06T11:06:43.638Z"},"it":{"title":"Il modello AI locale più veloce su Apple iPhone: Maple-Preview-20B-A1B debutta, con una velocità fino a 127 token/s","summary":"DeepGrove ha rilasciato il modello AI leggero Maple-Preview, che su iPhone raggiunge una velocità di 127 token/s, circa 13 volte quella del Bonsai 27B. Il modello ha un totale di 20,2 miliardi di parametri, utilizza un'architettura MoE, con 1,49 miliardi di parametri attivi, e propone uno schema di \"pesi trivalenti\" per ridurre l'uso della memoria. Attualmente è in fase di anteprima e l'uso della memoria durante l'elaborazione di 131.000 token è di soli 7,69 GB.","category":"模型更新","source":"IT之家（RSS）","aggregationSource":"IT之家（RSS）","pageTitle":"Il modello AI locale più veloce su Apple iPhone: Maple-Preview-20B-A1B debutta, con una velocità fino a 127 token/s - Aioga Notizie IA","description":"DeepGrove ha rilasciato il modello AI leggero Maple-Preview, che su iPhone raggiunge una velocità di 127 token/s, circa 13 volte quella del Bonsai 27B. Il modello ha un totale di 2...","url":"https://www.aioga.com/it/news/cmsh6fjwm002hronk4z70xx1v/","contentTranslated":true,"sourceHash":"b1d27ba2a743f55e","translatedAt":"2026-08-06T11:07:35.449Z"},"nl":{"title":"De snelste lokale AI-model op de Apple iPhone: Maple-Preview-20B-A1B gelanceerd, kan tot 127 tokens/s bereiken","summary":"DeepGrove heeft het lichtgewicht AI-model Maple-Preview uitgebracht, dat op de iPhone een snelheid van 127 tokens/s bereikt, ongeveer 13 keer die van Bonsai 27B. Het totale aantal parameters van dit model is 20,2 miljard, het maakt gebruik van een MoE-architectuur, met 1,49 miljard geactiveerde parameters, en introduceert een \"drie-waarde gewichts\"-oplossing om het geheugenverbruik te verlagen. Momenteel bevindt het zich in de previewfase, en bij het verwerken van 131.000 tokens bedraagt het geheugenverbruik slechts 7,69 GB.","category":"模型更新","source":"IT之家（RSS）","aggregationSource":"IT之家（RSS）","pageTitle":"De snelste lokale AI-model op de Apple iPhone: Maple-Preview-20B-A1B gelanceerd, kan tot 127 tokens/s bereiken - Aioga AI-nieuws","description":"DeepGrove heeft het lichtgewicht AI-model Maple-Preview uitgebracht, dat op de iPhone een snelheid van 127 tokens/s bereikt, ongeveer 13 keer die van Bonsai 27B. Het totale aantal...","url":"https://www.aioga.com/nl/news/cmsh6fjwm002hronk4z70xx1v/","contentTranslated":true,"sourceHash":"b1d27ba2a743f55e","translatedAt":"2026-08-06T11:07:34.265Z"},"tr":{"title":"Apple iPhone'daki en hızlı yerel AI modeli: Maple-Preview-20B-A1B tanıtıldı, 127 token/s hıza ulaşabiliyor","summary":"DeepGrove, iPhone üzerinde çalışma hızı saniyede 127 token’e ulaşan hafif bir AI modeli olan Maple-Preview’ı yayınladı, bu hız Bonsai 27B’nin yaklaşık 13 katı. Modelin toplam parametre sayısı 20,2 milyar, MoE mimarisi kullanıyor, etkin parametre sayısı 1,49 milyar ve bellek kullanımını azaltmak için \"üç değerli ağırlık\" çözümünü öneriyor. Şu anda ön izleme aşamasında olan model, 131.000 token işlenirken yalnızca 7,69 GB bellek kullanıyor.","category":"模型更新","source":"IT之家（RSS）","aggregationSource":"IT之家（RSS）","pageTitle":"Apple iPhone'daki en hızlı yerel AI modeli: Maple-Preview-20B-A1B tanıtıldı, 127 token/s hıza ulaşabiliyor - Aioga AI Haberleri","description":"DeepGrove, iPhone üzerinde çalışma hızı saniyede 127 token’e ulaşan hafif bir AI modeli olan Maple-Preview’ı yayınladı, bu hız Bonsai 27B’nin yaklaşık 13 katı. Modelin toplam param...","url":"https://www.aioga.com/tr/news/cmsh6fjwm002hronk4z70xx1v/","contentTranslated":true,"sourceHash":"b1d27ba2a743f55e","translatedAt":"2026-08-06T11:08:22.989Z"},"vi":{"title":"Mẫu AI cục bộ nhanh nhất trên iPhone của Apple: Maple-Preview-20B-A1B ra mắt, đạt tới 127 tokens/s","summary":"DeepGrove phát hành mô hình AI nhẹ Maple-Preview, chạy trên iPhone với tốc độ 127 tokens/s, khoảng gấp 13 lần Bonsai 27B. Tổng số tham số của mô hình này là 20,2 tỷ, sử dụng kiến trúc MoE, tham số kích hoạt 1,49 tỷ, và đề xuất phương án “trọng số ba giá trị” để giảm sử dụng bộ nhớ. Hiện tại đang ở giai đoạn xem trước, khi xử lý 131.000 tokens chỉ chiếm 7,69GB bộ nhớ.","category":"模型更新","source":"IT之家（RSS）","aggregationSource":"IT之家（RSS）","pageTitle":"Mẫu AI cục bộ nhanh nhất trên iPhone của Apple: Maple-Preview-20B-A1B ra mắt, đạt tới 127 tokens/s - Tin tức AI Aioga","description":"DeepGrove phát hành mô hình AI nhẹ Maple-Preview, chạy trên iPhone với tốc độ 127 tokens/s, khoảng gấp 13 lần Bonsai 27B. Tổng số tham số của mô hình này là 20,2 tỷ, sử dụng kiến t...","url":"https://www.aioga.com/vi/news/cmsh6fjwm002hronk4z70xx1v/","contentTranslated":true,"sourceHash":"b1d27ba2a743f55e","translatedAt":"2026-08-06T11:08:27.202Z"},"id":{"title":"Model AI lokal tercepat di iPhone Apple: Maple-Preview-20B-A1B hadir, bisa mencapai 127 token/detik","summary":"DeepGrove merilis model AI ringan Maple-Preview, dengan kecepatan menjalankan 127 token/detik di iPhone, sekitar 13 kali lipat dari Bonsai 27B. Model ini memiliki total 20,2 miliar parameter, menggunakan arsitektur MoE, dengan 1,49 miliar parameter aktif, dan mengusulkan solusi “bobot tiga nilai” untuk mengurangi penggunaan memori. Saat ini berada pada tahap pratinjau, penggunaan memori hanya 7,69GB saat memproses 131.000 token.","category":"模型更新","source":"IT之家（RSS）","aggregationSource":"IT之家（RSS）","pageTitle":"Model AI lokal tercepat di iPhone Apple: Maple-Preview-20B-A1B hadir, bisa mencapai 127 token/detik - Berita AI Aioga","description":"DeepGrove merilis model AI ringan Maple-Preview, dengan kecepatan menjalankan 127 token/detik di iPhone, sekitar 13 kali lipat dari Bonsai 27B. Model ini memiliki total 20,2 miliar...","url":"https://www.aioga.com/id/news/cmsh6fjwm002hronk4z70xx1v/","contentTranslated":true,"sourceHash":"b1d27ba2a743f55e","translatedAt":"2026-08-06T11:09:15.312Z"},"th":{"title":"โมเดล AI ในเครื่องที่เร็วที่สุดบน iPhone ของ Apple: Maple-Preview-20B-A1B เปิดตัวแล้ว ความเร็วสูงสุดถึง 127 tokens/s","summary":"DeepGrove เปิดตัวโมเดล AI ขนาดเบา Maple-Preview ทำงานบน iPhone ด้วยความเร็ว 127 tokens/s ซึ่งเร็วกว่า Bonsai 27B ประมาณ 13 เท่า โมเดลนี้มีจำนวนพารามิเตอร์ทั้งหมด 20.2 พันล้าน ใช้สถาปัตยกรรม MoE มีพารามิเตอร์ที่ถูกเปิดใช้งาน 1.49 พันล้าน และได้นำเสนอแนวทาง \"น้ำหนักสามค่า\" เพื่อลดการใช้หน่วยความจำ ปัจจุบันอยู่ในขั้นตอนพรีวิว เมื่อประมวลผล 131,000 tokens การใช้หน่วยความจำอยู่เพียง 7.69GB","category":"模型更新","source":"IT之家（RSS）","aggregationSource":"IT之家（RSS）","pageTitle":"โมเดล AI ในเครื่องที่เร็วที่สุดบน iPhone ของ Apple: Maple-Preview-20B-A1B เปิดตัวแล้ว ความเร็วสูงสุดถึง 127 tokens/s - ข่าว AI Aioga","description":"DeepGrove เปิดตัวโมเดล AI ขนาดเบา Maple-Preview ทำงานบน iPhone ด้วยความเร็ว 127 tokens/s ซึ่งเร็วกว่า Bonsai 27B ประมาณ 13 เท่า โมเดลนี้มีจำนวนพารามิเตอร์ทั้งหมด 20.2 พันล้าน ใช้สถ...","url":"https://www.aioga.com/th/news/cmsh6fjwm002hronk4z70xx1v/","contentTranslated":true,"sourceHash":"b1d27ba2a743f55e","translatedAt":"2026-08-06T11:09:25.389Z"},"pl":{"title":"Najszybszy lokalny model AI na iPhonie: Maple-Preview-20B-A1B wchodzi na rynek, osiągając 127 tokenów/s","summary":"DeepGrove wydało lekki model AI Maple-Preview, który działa na iPhone z prędkością 127 tokenów/s, co jest około 13 razy szybciej niż Bonsai 27B. Model ma łącznie 20,2 miliarda parametrów, wykorzystuje architekturę MoE, aktywne parametry wynoszą 1,49 miliarda, oraz wprowadza rozwiązanie \"trójwartościowych wag\" w celu zmniejszenia zużycia pamięci. Obecnie znajduje się w fazie podglądu, przy przetwarzaniu 131 000 tokenów zużycie pamięci wynosi tylko 7,69 GB.","category":"模型更新","source":"IT之家（RSS）","aggregationSource":"IT之家（RSS）","pageTitle":"Najszybszy lokalny model AI na iPhonie: Maple-Preview-20B-A1B wchodzi na rynek, osiągając 127 tokenów/s - Aioga Wiadomości AI","description":"DeepGrove wydało lekki model AI Maple-Preview, który działa na iPhone z prędkością 127 tokenów/s, co jest około 13 razy szybciej niż Bonsai 27B. Model ma łącznie 20,2 miliarda para...","url":"https://www.aioga.com/pl/news/cmsh6fjwm002hronk4z70xx1v/","contentTranslated":true,"sourceHash":"b1d27ba2a743f55e","translatedAt":"2026-08-06T11:10:24.638Z"}}}}