AI NEWSLatest AI NewsArticleTencent Hunyuan launches open-source flagship model Hy4preview with 770B total parameters and 1M contextPublished in Latest AI NewsTime :Aug 28, 2026Read :4minuteTencent Hunyuan officially released the new open-source flagship large model Hy4preview on August 28. The model has a total of 770B parameters and 49B activated parameters, with a context length of 1M, targeting real productivity scenarios. The model has been simultaneously open-sourced on platforms such as HuggingFace, Github, and Modelscope, and is available on Tencent Cloud TokenHub and OpenRouter. Users can directly experience it in products such as WorkBuddy/CodeBuddy, Yuanbao, and ima.In terms of capabilities, Hy4preview has made progress in four directions by leveraging high-quality data co-built with experts from Tencent’s internal fields such as software engineering, gaming, finance, and security: enhancing long-term development tasks’ understanding, planning, and debugging in software engineering; completing full delivery from data analysis to documents, tables, and presentations in office scenarios; supporting integration with Unreal5 and Unity engines via MCP for game development, generating playable demos from scratch through pure conversation; and covering areas such as molecular dynamics, condensed matter physics, and basic mathematics in scientific research.Internal blind test data shows that among 203 engineering tasks scored by 163 experts, Hy4preview achieved an average score of 2.99/4.00, slightly better than GLM-5.3 (2.92) and Kimi K3 (2.94). In scientific research breakthroughs, combined with the Hyra model, it advanced the volume lower bound of the century-old classic problem of three-dimensional Blaschke–Lebesgue from 0.380799 to 0.41104, leaving only about 2% distance from the final proof of the Meissner tetrahedron conjecture; and achieved a 2.0x speedup in simulating a 32, 512-atom phospholipid bilayer, reaching 54.9ms/step.Tencent stated that Hy4preview is an early version of the Hy4 iteration. Pre-training and post-training still have room for improvement, and there are known issues such as complex task long thinking. It will absorb real feedback through agile iteration. Its strategy of releasing an early version in an open-source manner continues the industry trend of shifting large model competition from parameter contests to productivity effectiveness verification.

AI NEWSLatest AI NewsArticleAnt Group Collaborates with CICC to Launch the First Financial-Enhanced Large Model Ling-3.0-flash-FinPublished in Latest AI NewsTime :Aug 28, 2026Read :2minuteAnt Group has recently partnered with China International Capital Corporation (CICC) and several industry experts to officially launch Ling-3.0-flash-Fin, the first financial-enhanced large model under Ant Bailing. Building upon the original architecture and parameter configuration, this model has significantly enhanced its core research and investment capabilities through extensive training on professional financial data.Regarding technical features and core capabilities, the model focuses on four key areas, including information retrieval. After undergoing multiple rigorous benchmark tests, it has demonstrated outstanding performance in specialized fields, while also improving its general capabilities. To promote the development of the industry ecosystem, FinFIRST, a financial search benchmark jointly developed by both parties, is about to be open-sourced.In terms of ecosystem and open plans, Ling-3.0-flash-Fin will offer one month of free API call services on the OpenRouter platform. The model weights are also planned to be officially open-sourced next week, bringing practical convenience and technical support to developers and financial institutions.

AI NEWSLatest AI NewsArticleTencent WorkBuddy Launches Integration with Hy4preview: Domestic and Overseas Versions Synchronized, Two-Week Free Trial StartedPublished in Latest AI NewsTime :Aug 28, 2026Read :3minuteTencent WorkBuddy announced on August 28 that it has launched the new open-source large model Hy4preview of Tencent Huan Yuan. The domestic and international versions were launched simultaneously, and a limited-time free policy was introduced: Hy4preview offers a two-week free trial, and the free usage period for the previous generation Hy3 has been extended by one month to September 30th.Hy4preview is the official open-source flagship model released by Tencent Huan Yuan. It has a total of 770B parameters and 49B activated parameters, with a context length of up to 1M. It is positioned for real productivity scenarios such as coding, office work, and science. It has been open-sourced on platforms such as HuggingFace, and is also available on Tencent Cloud TokenHub and OpenRouter.The domestic and international versions of WorkBuddy were launched simultaneously, allowing developers to directly call this model in programming and intelligent office workflows, achieving rapid integration from model release to product implementation.This combination strategy of initial access and limited-time free offers aims to attract developers to test the model’s capabilities and collect real feedback with zero barriers, while extending the free period of Hy3 to retain existing users. This continues Tencent Huan Yuan’s approach of promoting the accessibility of large models through open source and product ecosystem collaboration. The focus of competition among model manufacturers is shifting from parameter scale to productization and ecosystem service capabilities.

Join NowENLatest AI NewsTencent WorkBuddy Launches Integration with Hy4preview: Domestic and Overseas Versions Synchronized, Two-Week Free Trial StartedOn Aug 28, Tencent WorkBuddy first integrated Hunyuan open-source flagship model Hy4preview, launched globally. Hy4preview free for two weeks; Hy3 free extended to Sep 30. Total params 770B, active 49B, 1M context, targeting code, office, science productivity; open-sourced on HuggingFace etc…..1 days ago20.4KTencent Hunyuan Hy4 Preview Open Source: 770B Parameters, Million-Context, Standing at the Forefront of Open SourceTencent releases Hunyuan Hy4 preview open-source model: 770B total/49B active params, 1M context. Official: pretraining & post-training co-advance, top open-source tier. Live on HuggingFace, GitHub, ModelScope, Gitcode; integrated with Tencent Cloud TokenHub & OpenRouter. Blind tests beat GLM-5.3 & Kimi K3 in productivity…..1 days ago19.7KTencent Hunyuan Large Model Upgraded: 770B Flagship New Version Hy4preview Shockingly Open-Source, Directly Enters the First Tier of Open-Source ModelsTencent Hunyuan released the preview version of its flagship model, Hy4preview, with a total parameter count of 770B and an activated parameter count of 49B. It supports a context length of 1M. This model performs exceptionally well in real productivity tasks such as coding, office work, and science, ranking among the first-tier open-source models. Its comprehensive performance in internal blind tests exceeds that of similar competitors, and it has already been applied in multiple fields within Tencent.1 days ago15.4KWeibo Reports that the Chat Records Between Sun Yuchong and Eileen Gu Were Fabricated by AI and the Rumor-Mongering Accounts Have Been Permanently ClosedWeibo announced that the chat records between Sun Yuchong and Eileen Gu, which were circulated online, were fabricated rumors, and relevant content has been handled. Sun Yuchong’s ex-girlfriend Zeng Ying denied posting nude photos, stating that the entire set of records was generated by AI, and exposed that the rumor-mongers bought traffic to push hot topics, showing obvious traces of black industry involvement, with the involved accounts only following one person across the platform.1 days ago17.3KAnt Group Collaborates with CICC to Launch the First Financial-Enhanced Large Model Ling-3.0-flash-FinAnt Group, CICC and partners released Ling-3.0-flash-Fin, Ant Bailing’s first finance-enhanced LLM. It retains the original architecture, trained on massive financial data, strengthening investment research and four core capabilities like information retrieval, with strong professional and general performance…..1 days ago17.4KTencent Hunyuan launches open-source flagship model Hy4preview with 770B total parameters and 1M contextTencent Hunyuan released the open-source flagship large model Hy4preview on August 28th, with a total of 770B parameters and 49B activated parameters, and a context length of 1M, positioning it as real productivity. The model has been open-sourced on HuggingFace, GitHub, ModelScope, and is also available on Tencent Cloud TokenHub and OpenRouter. It can be experienced through WorkBuddy/CodeBuddy, Yaobao, and IMA. It was built using expert data from fields such as software engineering, games, finance, and security.1 days ago18.8KAnthropic Launches Major Model Hardware Standard MHS, AI Agent Officially Advances to Physical ControlAnthropic introduced the Model Hardware Standard (MHS), aiming to enable large model-driven AI agents to safely and efficiently control physical devices, marking a crucial step in bridging the virtual and real worlds. The standard originated from collaboration with the Howard Hughes Medical Institute’s Janelia Campus, and is now available for preview to early research laboratories and advanced manufacturers, with significant improvements in practical application efficiency.1 days ago15.4KAnthropic Launches MHS Hardware Standard, Making Its Strong Entry into the Physical AI ArenaOn Aug 27, Anthropic launched the Model Hardware Standard (MHS) as a research preview, seen as its first public move into embodied AI. Like MCP, it offers unified control and communication, giving AI a ‘body’ to operate autonomous vehicles, robotic arms, perceive, act, and learn…..1 days ago15.7KAlibaba Launches New Qoder: Upgraded from AI Programming Tool to Intelligent Agent Workstation for EveryoneAlibaba launched Qoder, upgrading its AI coding IDE into an agent workbench centered on coding. Users state goals in natural language; it understands context, plans, calls tools, executes and verifies, with visibility, adjustment or takeover throughout. In one year, over 6M users and 100K enterprises. The new version reflects three collaboration paradigm shifts…..1 days ago17.5KMidjourney V8.2 Launches Image Editing Model with Full Function Support for Instruction Fine-tuning and Canvas ExpansionMidjourney recently beta-tested its first V8.2 image editing model, break

Join NowENPremium Membership ·Limited-Time Pricing— Save TodayChoose a monthly plan that fits your needs. Each tier includes tailored points allowances and flexible usage limits for confident monitoring.Your current plan isFreeEssentialStart with essential GEO monitoring and analysis$14.9/30 daysGet PremiumPoints IncludedOne-time points 7,000Daily bonus points 100AI Query Mining50 usesGEO Rank Check50 uses5 AI platforms:GEO Brand Score Audit3 uses5 AI platforms:GEO Link Citation Tracking5 uses5 AI platforms:GEO Rank TrackingNot includedData Export：Not includedReport Generation：Not included7-Day Featured Slot in AI DirectoryNot includedDedicated SupportNot includedMost PopularGrowthBuilt for growing brands and daily GEO operations$45/30 daysGet PremiumPoints IncludedOne-time points 25,000Daily bonus points 200AI Query Mining150 usesGEO Rank Check100 uses7 AI platforms:GEO Brand Score Audit8 uses7 AI platforms:GEO Link Citation Tracking15 uses7 AI platforms:GEO Rank TrackingNot includedData Export：Not includedReport Generation：Not included7-Day Featured Slot in AI DirectoryNot includedDedicated SupportNot includedProProfessional GEO monitoring and continuous growth analysis$109/30 daysGet PremiumPoints IncludedOne-time points 65,000Daily bonus points 300AI Query Mining300 usesGEO Rank Check300 uses10 AI platforms:GEO Brand Score Audit20 uses10 AI platforms:GEO Link Citation Tracking30 uses10 AI platforms:GEO Rank TrackingUnlimited10 AI platforms:Data Export：IncludedReport Generation：Included7-Day Featured Slot in AI DirectoryNot includedDedicated SupportNot includedEnterprise PlanEnterpriseScalable GEO operations for enterprises and agencies$199/30 daysGet PremiumPoints IncludedOne-time points 120,000Daily bonus points 500AI Query Mining800 usesGEO Rank Check600 uses10 AI platforms:GEO Brand Score Audit30 uses10 AI platforms:GEO Link Citation Tracking60 uses10 AI platforms:GEO Rank TrackingUnlimited10 AI platforms:Data Export：IncludedReport Generation：Included7-Day Featured Slot in AI Directory1 usesDedicated SupportIncludedCompare PlansCompare usage allowances and premium benefits across plans at a glanceFree$0Get PremiumEssential$99/30 daysGet PremiumGrowth$299/30 daysGet PremiumPro$699/30 daysGet PremiumEnterprise$1,299/30 daysGet PremiumPoints IncludedOne-time points07,00025,00065,000120,000Daily bonus points100100200300500AI Query MiningQuery allowance550150300800GEO Rank CheckQuery allowance350100300600AI platformsGEO Brand Score AuditQuery allowance1382030AI platformsGEO Link Citation TrackingQuery allowance15153060AI platformsGEO Rank TrackingGEO Rank TrackingNot includedNot includedNot includedUnlimitedUnlimitedData ExportNot includedNot includedNot includedIncludedIncludedReport GenerationNot includedNot includedNot includedIncludedIncluded7-Day Featured Slot in AI Directory7-Day Featured Slot in AI DirectoryNot includedNot includedNot includedNot included1Dedicated SupportDedicated SupportNot includedNot includedNot includedNot includedIncludedFrequently Asked QuestionsCommon questions about Aibase memberships, credits, GEO monitoring, and billingWhat is the difference between monthly membership credits and daily login bonus credits?Monthly membership credits are the core AI computing allowance included with your membership and are issued once per membership benefit cycle.Daily login bonus credits are an additional benefit. They are automatically issued when a member first signs in to or visits Aibase on a given day, with no manual claim required.Daily bonus credits are valid only for the day they are issued and do not roll over to the next day.Do I need to manually claim daily bonus credits? Are missed days credited later?No manual claim is required. Daily bonus credits are automatically issued when you first sign in to or visit Aibase that day.If you do not sign in to or visit Aibase on a given day, that day’s bonus credits will not be issued and will not be credited retroactively.Daily bonus credits are an additional membership benefit and do not affect your regular monthly membership credits.Do monthly membership credits roll over to the next month?No. Membership credits are issued in 30-day benefit cycles.Unused membership credits expire at the end of the current benefit cycle and do not roll over into the next cycle.At the beginning of a new benefit cycle, the credit allowance for your current membership plan is issued again.If I buy a 3-, 6-, or 9-month membership, will all credits and usage quotas be issued at once?No. A multi-month purchase provides membership access for the selected duration; it does not issue all future credits and usage quotas upfront.Membership benefits are still managed in 30-day cycles. Credits are issued and tool usage quotas are reset at the start of each cycle.For example, a 3-month membership consists of three consecutive 30-day benefit cycles.In what order are different credit balances used?Credits are deducted in the following order: daily login bonus credits →

Gemini 3.5 Transcribe – 谷歌推出的最新语音转文本模型AI工具3天前发布AI小集02Gemini 3.5 Transcribe是什么Gemini 3.5 Transcribe 是 Google 推出的最新语音转文本模型，支持实时流式和预录音频处理两种模式，前者通过 Live API 实现亚秒级延迟，后者支持说话人归属与词级时间戳。Gemini 3.5 Transcribe 能自动清理填充词与自我纠正、识别 85 种以上语言及实时语言切换、自定义专业词汇，词错误率低至 2.6%，支持调用其他 Gemini 模型完成图像生成等复杂任务。Gemini 3.5 Transcribe的主要功能实时流式转录：通过 Live API 提供亚秒级延迟的连续双向语音转文本服务。预录音频转写：通过 Interactions API 为录音文件提供带说话人归属和词级时间戳的精确转录。智能文本清理：自动去除”嗯””啊”等填充词，修正自我纠正语句，并输出格式化文本。自定义词汇适配：根据用户提供的专业术语和特殊拼写调整转录结果，准确识别邮政编码、订单号等字母数字组合。多语言自动检测：支持 85 种以上语言的自动识别与转录，并能处理实时语言切换和多种口音方言。多说话人区分：在预录音频中准确归属最多三位说话人的发言内容。跨模型功能调用：在转录过程中可调用其他 Gemini 模型执行图像生成、文件分析等复杂任务。屏幕上下文融合：结合设备屏幕内容和对话历史提升特定场景下的转录准确度。Gemini 3.5 Transcribe的技术原理端到端音频理解：模型直接从原始音频波形生成文本，避免中间表示引入的误差累积，同时保留语音中的语调、停顿等副语言信息用于语义消歧。原生多模态融合：基于 Gemini 3.5 统一架构，音频与文本在共享的 Transformer 注意力空间中进行联合编码，使模型能建立声学特征与语义概念的直接映射。上下文感知推理：通过长上下文窗口同时处理语音信号、屏幕截图文本和对话历史，用交叉注意力机制动态加权相关上下文，提升专有名词、文件名称和领域术语的识别精度。增量流式解码：采用自适应分块与投机解码策略，在音频输入持续到达的同时进行局部语义整合，平衡低延迟需求与输出稳定性，支持双向交互中的即时响应与打断处理。如何使用Gemini 3.5 Transcribe开发者接入：通过 Gemini API 在 Google AI Studio 或 Gemini Enterprise Agent Platform 调用模型。实时语音交互：用 Live API 调用gemini-3.5-transcribe-live实现双向流式转录。录音文件处理：用 Interactions API 调用gemini-3.5-transcribe进行带时间戳的离线转写。macOS 桌面端：在 Gemini macOS 应用中直接用语音输入、编辑文本并调用其他模型功能。Android 输入法：在 Gboard 中开启 Rambler 功能，语音输入自动清理并支持语音改稿。Chrome 浏览器：即将支持在任意网页输入框中语音输入与打字。企业部署：通过 Gemini Enterprise Agent Platform 集成至客服与内部工作流。第三方框架：借助 Agora、LiveKit、Vercel 等已接入 Live API 的平台快速搭建语音应用。Gemini 3.5 Transcribe的核心优势转录精度领先：流式场景词错误率低至 4.0%，非流式低至 2.6%，在嘈杂环境中仍能准确捕获字母数字实体。延迟大幅优化：相比前代 Chirp 3，最终转录交付时间缩短 70%，实时流式响应达到亚秒级。智能语义清理：自动去除填充词、修正自我纠正语句，并输出可直接使用的格式化文本。多语言无缝切换：自动识别并转录 85 种以上语言，支持对话中的实时语言切换与多方言口音适配。说话人精准归属：预录音频支持最多三位说话人的发言区分，并附带词级时间戳便于回溯定位。屏幕上下文感知：结合设备屏幕内容与对话历史，显著提升文件名称、专业术语和活跃文档的识别准确度。跨模型任务协同：内置 Function Calling 能力，可在转录过程中调用其他 Gemini 模型完成图像生成、文件分析等复杂操作。Gemini 3.5 Transcribe的项目地址项目官网：https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-5-transcribe/Gemini 3.5 Transcribe的同类竞品对比对比维度Gemini 3.5 TranscribeOpenAI GPT-4o-transcribe定位智能语音转录模型，支持实时流式与预录音频GPT-4o 架构的专用转录模型，API 托管服务词错误率 WER流式 4.0% / 非流式 2.6%约 2.5%（干净音频，行业领先）实时流式延迟亚秒级，原生双向连续流式支持流式，通过 Realtime API 实现智能文本清理自动去除填充词、修正自我纠正、格式化输出不支持，输出原始口语化文本说话人归属原生支持最多 3 人区分 + 词级时间戳不支持原生，需额外调用 diarize 端点Gemini 3.5 Transcribe的应用场景实时会议智能纪要：在多人会议中实时流式转录，自动区分说话人、清理口语填充词，调用 Gemini 模型即时生成会议摘要与待办事项。多语言客服语音助手：通过 Live API 构建亚秒级延迟的语音客服系统，自动识别客户语言并实时切换，结合自定义词汇准确识别订单号、产品型号等关键信息。医疗与法律口述记录：医生或律师通过语音口述病历、庭审记录，模型自动清理自我纠正语句并格式化专业术语，输出可直接归档的标准文档。跨语言实时字幕与翻译：为直播、视频会议提供多语言实时字幕，用 85+ 语言自动检测能力，结合实时语言切换实现无缝跨语言沟通。语音驱动编程与办公自动化：在 Google AI Studio 或 Gemini macOS 应用中，开发者通过语音描述需求即可生成代码，办公人员语音指令调用屏幕上下文完成跨应用文件总结与图像生成。# AI工具# AI项目和框架©版权声明本站文章版权归AI工具集所有，未经允许禁止任何形式的转载。上一篇Hy-MT2-1.8B - 腾讯混元推出的端侧翻译大模型下一篇怎么用 MetaDig 做 AI 短视频，附教程实测案例相关文章AlphaClaw – 熵简科技推出的金融投研AI AgentAI小集4Silimini – AI动态照片应用，静态照片转换成生动的动态表情AI小集2灵羽助手 – AI桌面助手，支持微信、浏览器、VSCode、PDF等多种应用AI小集2Model1 – DeepSeek代码库更新的新模型版本AI小集2Mercury – Inception Labs推出的扩散语言模型AI小集3OpenNof1 – 开源的AI自主交易系统，实时交易监控AI小集2暂无评论再想想发表评论暂无评论…热门工具Loomy即梦AISekoAiPPT秘塔AI搜索妙呀堆友Agent美图设计室绘蛙AIupdream办公小浣熊秒哒最新收录AlphaVowXCellMetaDigChatArtLexcat造剧最新文章秒悟Meoo – 1分钟生成前后端完整的网站，真0门槛开发！39秒前Rome – 开源的智能体操作系统，自然语言生成完整应用1分钟前Headlong – Laude Institute 开源的 AI Agent 微框架3分钟前Miya – 时域科技推出的 AI 音乐应用，画面可随歌词律动1天前Zing-0.5 – Loopit 开源的实时交互视频世界模型1天前2026 年 AI 生成图文店门头效果图 – 7 款工具深度横评2天前Hy4 preview – 腾讯混元开源的新一代旗舰大模型2天前Gemini Omni 1.1 Flash – 谷歌推出的 AI 视频生成模型2天前Twoo – 双人共用 AI 伙伴应用，能同时听懂双方对话2天前Prime Agent – Prime Intellect开源的长周期AI Agent运行框架2天前shuohao-skills – 开源AI短剧制作Skill集合，包含完整工作流2天前怎么用 MetaDig 做 AI 短视频，附教程实测案例3天前Hy-MT2-1.8B – 腾讯混元推出的端侧翻译大模型3天前GLM-5.3-Flash – 智谱开源的原生多模态模型，即Ox Alpha3天前Qwen3.8-Flash – 阿里通义推出的多模态 MoE 模型3天前

感谢IT之家网友华南吴彦祖、乌蝇哥的左手的线索投递！IT之家8 月 27 日消息，在外界仍等待 Gemini 3.5 Pro 发布之际，谷歌先推出了 3.5 系列中的另一款模型。该公司今日发布了 Gemini 3.5 Transcribe 语音转文本模型，这是一款旨在优化语音输入体验的 AI 模型，能够自动删除语音中的“嗯”“呃”等口头禅以及说错后自行纠正的内容，最终生成更加通顺、整洁的文本。IT之家注意到，这款模型目前已经为 Pixel 11 上 Gboard 键盘的“Rambler”功能提供支持，接下来还将陆续进入谷歌旗下更多产品和服务。谷歌表示，与此前名为 Chirp 3 的语音转文字引擎相比，Gemini 3.5 Transcribe 在速度和准确性方面都有提升。从用户说话到最终生成转录文本，整体速度预计可提升约 70%；在实时语音场景下，其错误率已经降至 5.5%。不过，准确率的提升幅度并不算特别大。按照谷歌的数据，Chirp 3 的错误率为 7.32%。但对于经常使用语音输入的用户来说，即便只是少打几个错字，也意味着后续需要手动修改的内容更少，因此依然是一项实用的改进。Gemini 3.5 Transcribe 的能力也不只是“听清楚”用户说了什么，它还试图理解用户真正想表达的意思。在说话过程中，模型可以自动删除“嗯”“呃”等影响语句流畅度的口头禅。如果用户说错话后立即进行修正，它也能够实时调整已经生成的文本。此外，用户还可以提供自定义词汇表，让模型更好地识别专业术语和特定领域的词汇。该模型目前支持 85 种语言，并且可以处理最多包含 3 名说话者的预录音频。当然，这种功能也存在一个明显的问题：用户需要相信 AI 能够准确理解自己真正想表达的内容。至少在较短的文本中，从 Rambler 功能的实际体验来看，模型确实能够较好地清理前后不一致的表达和说话时的磕绊，让文本看起来更加自然。但与此同时，Gemini 3.5 Transcribe 并不只是简单地把语音转换成文字。AI 在整理内容的过程中，实际上可能会改变用户原本的措辞。因此，在一些需要严格保留原话、不能随意修改表述的场景中，这种“智能润色”未必适合使用。

Hy-MT2-1.8B – 腾讯混元推出的端侧翻译大模型AI工具3天前发布AI小集02Hy-MT2-1.8B是什么Hy-MT2-1.8B 是腾讯混元推出的端侧翻译大模型，仅 1.8B 参数却支持 33 个语种互译，翻译质量在 FLORES-200 等基准上超越微软、豆包等商业 API。模型通过 2-bit（574MB）与 1.25-bit（440MB）极致量化，可在手机、PC 端本地流畅运行，已落地哔哩哔哩直播弹幕实时翻译，实现低成本、高隐私的端侧高质量翻译。Hy-MT2-1.8B的主要功能33 语种互译：覆盖中英日韩等主流语言及少数民族语言，支持任意方向互译。端侧本地运行：通过极低比特量化将模型压缩至 440574MB，普通手机与 PC 均可常驻后台。实时翻译：单条弹幕 500800ms 响应，已接入哔哩哔哩直播弹幕等高频场景。隐私保护：翻译全程在设备本地完成，无需上传云端，数据不出端。Hy-MT2-1.8B的技术原理2-bit 极致量化：采用拉伸弹性量化（SEQ）将参数量化至 {-1.5, -0.5, 0.5, 1.5} 四个离散值，结合量化感知蒸馏（QAD）在压缩至 574MB 的同时保持几乎无损的翻译质量。1.25-bit 稀疏量化（Sherry）：混元自研的 ACL 2026 Oral 方案，核心为”细粒度稀疏”策略——每 4 个参数中保留 3 个 ±1、1 个置 0，平均每个参数仅需 1.25-bit；配合专为 CPU 设计的 STQ 内核，模型体积降至 440MB，可常驻后台运行。x86 算子适配优化：英特尔团队针对低比特格式优化向量化、权重重排与 VNNI 指令融合，使 Q2_0C 与 STQ 在主流酷睿处理器上的 token 吞吐分别提升约 2.7 倍和 5 倍，将极低 bit 方案从移动端扩展到 PC 与边缘设备。微信关注回复“开源”，加入AI开源项目交流群如何使用Hy-MT2-1.8B下载模型权重：访问 Hugging Face 仓库huggingface.co/collections/AngelSlim/hy-low-bit-model，下载 2-bit 或 1.25-bit 量化后的模型文件。获取推理框架：从 GitHub 克隆github.com/tencent/AngelSlim完整代码，该仓库包含极低比特量化的推理内核与加载逻辑。加载与推理：用 AngelSlim 框架加载下载的权重，在本地 CPU 上执行翻译推理；x86 设备可调用英特尔优化后的 VNNI 指令集内核以获得最佳性能。Hy-MT2-1.8B的核心优势体积极小质量极高：1.8B 参数经 2-bit / 1.25-bit 量化后仅 440574MB，翻译质量却超越微软、豆包等商业 API。端侧隐私优先：翻译全程在设备本地完成，无需联网上传，彻底杜绝数据泄露风险。跨平台无缝运行：同时适配 ARM 移动端与 x86 PC 端，经英特尔算子优化后在酷睿处理器上吞吐提升 2.75 倍。真实场景已落地：已接入哔哩哔哩直播弹幕实时翻译，单条 500~800ms 响应，能准确处理网络流行语。多语种全覆盖：原生支持 33 个语种互译，涵盖主流语言及少数民族语言，满足全球化沟通需求。Hy-MT2-1.8B的项目地址GitHub仓库：https://github.com/Tencent-Hunyuan/Hy-MT2HuggingFace模型库：https://huggingface.co/collections/AngelSlim/hy-low-bit-modelHy-MT2-1.8B的同类竞品对比对比维度Hy-MT2-1.8BGoogleTranslateGemma-4B参数规模1.8B4B支持语种33 个语种互译 + 5 种民族语言/方言覆盖主流语言对端侧体积440MB(1.25-bit) / 574MB (2-bit)2GB+ (原生 FP16，需自行量化压缩)核心量化技术Sherry 1.25-bit 细粒度稀疏量化（ACL 2026 Oral）无官方量化方案，需依赖第三方工具翻译质量FLORES-200 整体优于微软、豆包等商业 API接近研究级水平，4B 版为移动端效果妥协版推理框架AngelSlim（自研，含 x86 VNNI 优化内核）需借助 TensorFlow Lite、ExecuTorch 等通用框架商业落地已接入哔哩哔哩直播弹幕实时翻译主要用于研究/开发者实验，无大规模商业集成开源协议权重+代码全开源Apache 2.0 开源，可自由商用Hy-MT2-1.8B的应用场景直播弹幕实时翻译：已落地哔哩哔哩，单条 500800ms 响应，可准确翻译网络流行语与梗。离线即时通讯：集成于社交 App，在无网络环境下实现跨国聊天实时互译，保障对话隐私。移动端网页/文档翻译：作为浏览器插件或系统级翻译服务，本地完成外文网页、PDF 文档翻译。游戏内实时本地化：嵌入手游或 PC 游戏客户端，实时翻译外服游戏聊天、任务文本与 UI。跨境电商内容翻译：在商家端本地批量翻译商品标题、详情页与客服话术，降低多语言运营成本。# AI工具# AI项目和框架©版权声明本站文章版权归AI工具集所有，未经允许禁止任何形式的转载。上一篇GLM-5.3-Flash - 智谱开源的原生多模态模型，即Ox Alpha下一篇Gemini 3.5 Transcribe - 谷歌推出的最新语音转文本模型相关文章LineArt – 吉林大学等机构推出的设计绘图外观迁移框架AI小集3Skywork R1V4-Lite – 昆仑万维推出的轻量级多模态智能体AI小集2ChatTTSPlus – 开源文本转语音工具，ChatTTS扩展版支持语音克隆AI小集2Mellum – JetBrains 推出专为开发者设计的大语言模型AI小集3AnuNeko – AI聊天工具，体验人性化和情绪化对话体验AI小集9TurboDiffusion – 清华大学等推出的视频生成加速框架AI小集4暂无评论再想想发表评论暂无评论…热门工具Loomy即梦AISekoAiPPT秘塔AI搜索妙呀堆友Agent美图设计室绘蛙AIupdream办公小浣熊秒哒最新收录AlphaVowXCellMetaDigChatArtLexcat造剧最新文章秒悟Meoo – 1分钟生成前后端完整的网站，真0门槛开发！39秒前Rome – 开源的智能体操作系统，自然语言生成完整应用1分钟前Headlong – Laude Institute 开源的 AI Agent 微框架3分钟前Miya – 时域科技推出的 AI 音乐应用，画面可随歌词律动1天前Zing-0.5 – Loopit 开源的实时交互视频世界模型1天前2026 年 AI 生成图文店门头效果图 – 7 款工具深度横评2天前Hy4 preview – 腾讯混元开源的新一代旗舰大模型2天前Gemini Omni 1.1 Flash – 谷歌推出的 AI 视频生成模型2天前Twoo – 双人共用 AI 伙伴应用，能同时听懂双方对话2天前Prime Agent – Prime Intellect开源的长周期AI Agent运行框架2天前shuohao-skills – 开源AI短剧制作Skill集合，包含完整工作流2天前怎么用 MetaDig 做 AI 短视频，附教程实测案例3天前Gemini 3.5 Transcribe – 谷歌推出的最新语音转文本模型3天前GLM-5.3-Flash – 智谱开源的原生多模态模型，即Ox Alpha3天前Qwen3.8-Flash – 阿里通义推出的多模态 MoE 模型3天前

Qwen3.8-Flash – 阿里通义推出的多模态 MoE 模型AI工具3天前更新AI小集03Qwen3.8-Flash是什么Qwen3.8-Flash 是阿里云通义千问推出的多模态 MoE 模型，总参数量 125B，每 token 仅激活 6B，原生支持 262K 上下文。模型采用 GDN+QSA 混合注意力、门控残差、N-gram 嵌入及 Muon 优化器四大架构创新，训练成本仅为 Qwen3.7-Plus 的 1/9，在编码和办公任务上能力更强。Qwen3.8-Flash的主要功能超长上下文处理：原生支持 262K tokens，可扩展至 1M，胜任长文档与代码库分析。代码智能体：支持端到端软件工程任务，包括代码生成、调试与多语言开发。办公自动化：执行长程办公流程与专业工作任务，集成工具调用能力。多模态理解：理解长视频、科学图表、视觉数学问题及现实世界场景。设备与界面操控：通过视觉感知操控网页、Android 设备及桌面操作系统。推理与问答：覆盖科学推理、竞技编程、指令遵循与多学科综合推理。Qwen3.8-Flash的技术原理GDN + QSA Hybrid Attention：模型在每四层中安排三层 Gated DeltaNet 和一层 Qwen Sparse Attention，实现”高效记忆 + 精准查找”的混合注意力机制，在 1M 上下文下 Prefill 吞吐达到前代的 8.6 倍。Gated Residual（GR）：将传统单一残差流扩展为四条并行分支，通过动态门控控制各层信息的读写比例，保留长距离信息通道，抑制激活异常值，同时支持 FP8 存储大幅降低访存开销。N-gram Embedding：在常规词嵌入之外，引入 51B 参数的 N-gram 嵌入层，结合局部上下文查表捕捉短语搭配模式；该层可卸载至 Host Memory 并通过异步预取与计算重叠，几乎不增加 token 级计算量。Muon Optimizer：针对主要线性映射权重采用 Muon 正交化优化，对 Embedding 和 Router 等保留 AdamW；重新拟合 Scaling Law 后支持更大学习率与批次，取消无效的 Batch Size Warmup 策略。微信关注回复“开源”，加入AI开源项目交流群如何使用Qwen3.8-Flash开源权重部署：从 Hugging Face 下载 Qwen3.8-Flash-Next 权重，在本地或云端进行推理部署。API 服务调用：通过千问AI平台调用官方 API，输入 1 元/百万 tokens、输出 3 元/百万 tokens。千问办公使用：在”千问办公”中直接体验，用于长文档处理、代码生成与办公自动化任务。查阅技术报告：访问 GitHub 仓库中的技术报告，了解架构细节与训练方法以进行深度定制。Qwen3.8-Flash的核心优势极致成本效率：训练开销仅为 Qwen3.7-Plus 的 1/9，API 定价输入 1 元/百万 tokens，用极低算力成本实现旗舰级性能。超长上下文领先：原生 262K 可扩展至 1M tokens，1M 上下文下 Prefill 吞吐达到前代的 8.6 倍，长序列处理速度与精度兼具。Agent 与编码能力突出：在代码智能体、软件工程、长程办公任务等多项基准上显著超越同级模型，胜任复杂生产级开发工作。下一代架构创新：采用 GDN+QSA 混合注意力、门控残差与 N-gram 嵌入等全新设计，是 Qwen4 系列架构的提前验证版。多模态操控能力：支持长视频理解、视觉网页开发、Android 及桌面设备操控，实现真正的多模态智能体交互。开源与工程友好：Flash-Next 权重已开源，GR 支持 FP8、N-gram 嵌入可卸载至 Host Memory，降低部署门槛与显存占用。Qwen3.8-Flash的项目地址项目官网：https://qwen.ai/blog?id=qwen3.8-flash-nextHuggingFace模型库：https://huggingface.co/Qwen/Qwen3.8-Flash-Next?spm=a2ty_o06.30285417.0.0.1d73c921FsyOPe&file=Qwen3.8-Flash-Next技术论文：https://github.com/QwenLM/Qwen3.8-Flash-Next/blob/main/tech_report.pdfQwen3.8-Flash的同类竞品对比对比维度Qwen3.8-Flash-NextDeepSeek-V4-FlashClaude-Opus-4.6总参数量125B + 51B N-gram284B—激活参数6B13B—Agentic Coding(DeepSWE 1.1)58.754.4—SWE-bench Pro62.556.053.4多语言软件工程81.0—77.5长程办公(CoWorkBench)73.945.168.2专业工作(JobBench)55.741.336.6科学推理(GPQA Diamond)91.790.891.3竞技编程(LiveCodeBench)91.990.688.8指令遵循(IFBench)81.379.262.5多模态工具(ClawEval-MM)64.4/ 60.4—52.5 / 54.7移动操控(AndroidWorld)84.5—62.0计算机使用(OSWorld 2.0)19.4 /52.3——视觉网页开发(Vision2Web)64.0——具身智能(ERQA)72.3—40.8长视频理解(LVBench)76.6—63.0视觉数学(MathVision w/ CI)95.7—65.5科学图表分析(CharXiv w/ CI)90.6—66.0Qwen3.8-Flash的应用场超长文档与知识库智能分析：用 1M tokens 超长上下文，对整本技术手册、法律合同或大型代码库进行一次性深度理解与问答，无需分段处理。端到端软件开发与代码维护：作为代码智能体（Agentic Coding）完成需求分析、代码生成、Bug 修复、单元测试编写及多语言项目维护，覆盖 SWE-bench 级别复杂任务。企业办公流程自动化：执行跨平台长程办公任务，如自动整理会议纪要、批量处理 Excel 报表、撰写并排版商业文档，集成工具调用实现端到端闭环。多模态内容理解与教育辅导：解析长视频课程、科学论文图表、视觉数学题目，提供步骤化讲解与答案验证，支持 STEM 领域的深度教学辅助。跨设备智能体操控：通过视觉感知与推理能力，自动操控 Android 手机完成 App 测试、在桌面操作系统中执行文件管理，或根据设计稿自动生成可部署的网页前端。# AI工具# AI项目和框架©版权声明本站文章版权归AI工具集所有，未经允许禁止任何形式的转载。上一篇Omarchy - 开源的AI原生Linux桌面 ，自然语言生成桌面主题下一篇GLM-5.3-Flash - 智谱开源的原生多模态模型，即Ox Alpha相关文章Ming-omni-tts – 蚂蚁集团开源的统一音频生成模型AI小集2video-use – Browser Use 团队开源的 AI 视频剪辑 AgentAI小集2PxBee – 免费AI图片编辑平台，自动识别移除背景AI小集3Cua – 专为 MacOS 推出的开源 AI Agent项目AI小集4Gemini Robotics On-Device – 谷歌推出的首个本地具身智能模型AI小集3DeepEyesV2 – 小红书开源的多模态智能体模型AI小集2暂无评论再想想发表评论暂无评论…热门工具Loomy即梦AISekoAiPPT秘塔AI搜索妙呀堆友Agent美图设计室绘蛙AIupdream办公小浣熊秒哒最新收录AlphaVowXCellMetaDigChatArtLexcat造剧最新文章秒悟Meoo – 1分钟生成前后端完整的网站，真0门槛开发！14分钟前Rome – 开源的智能体操作系统，自然语言生成完整应用14分钟前Headlong – Laude Institute 开源的 AI Agent 微框架16分钟前Miya – 时域科技推出的 AI 音乐应用，画面可随歌词律动1天前Zing-0.5 – Loopit 开源的实时交互视频世界模型1天前2026 年 AI 生成图文店门头效果图 – 7 款工具深度横评2天前Hy4 preview – 腾讯混元开源的新一代旗舰大模型2天前Gemini Omni 1.1 Flash – 谷歌推出的 AI 视频生成模型2天前Twoo – 双人共用 AI 伙伴应用，能同时听懂双方对话2天前Prime Agent – Prime Intellect开源的长周期AI Agent运行框架2天前shuohao-skills – 开源AI短剧制作Skill集合，包含完整工作流2天前怎么用 MetaDig 做 AI 短视频，附教程实测案例3天前Gemini 3.5 Transcribe – 谷歌推出的最新语音转文本模型3天前Hy-MT2-1.8B – 腾讯混元推出的端侧翻译大模型3天前GLM-5.3-Flash – 智谱开源的原生多模态模型，即Ox Alpha3天前

GLM-5.3-Flash – 智谱开源的原生多模态模型，即Ox AlphaAI工具3天前更新AI小集03GLM-5.3-Flash是什么GLM-5.3-Flash 是智谱最新推出的开源大模型，总参数三百二十亿，激活参数仅十八亿，为 GLM-5 系列首个原生多模态模型。模型综合智能指数达五十七分，与Claude Opus 4.8持平，定价仅为其四十分之一。模型采用稀疏与线性注意力混合架构及流形约束超连接技术，支持视觉编码、代码生成和专业文档处理，首次实现国产芯片集群大规模高效部署，端到端性能提升三倍。GLM-5.3-Flash的主要功能原生多模态理解：支持文本与图像的联合理解与生成，是 GLM-5 系列首个原生多模态模型。视觉编码：将视觉能力原生集成进模型，使其能自主判断何时需要”观察”，并利用视觉反馈指导编程与行动。代码生成与迭代：支持前端、游戏、3D 仿真等开发任务，能根据渲染和交互反馈自我检验并迭代改进代码。专业文档处理：针对 PPTX、PDF、DOCX、XLSX 等 Office 文档优化，具备审美判断与自我验证能力。金融与法律工作：覆盖金融研报生成、建模分析，以及合同审查、法律文书起草等专业场景。Agent 协同：通过 Browser Use Agent 与 Computer Use Agent 在代码、浏览器和图形界面间协同操作。GLM-5.3-Flash的技术原理混合注意力架构：GLM-5.3-Flash 采用线性注意力与稀疏注意力混合架构，线性注意力通过递归机制捕获局部依赖，稀疏注意力借助轻量级索引器召回全局上下文；同时引入 IndexPool 将索引器缓存向量从 4 个压缩至 1 个，使注意力计算量与 KV 缓存较 GLM-5.3 分别降低 3.01 倍和 4.44 倍，在保持精准长上下文能力的同时大幅降低服务成本。流形约束超连接（mHC）：模型采用 mHC 技术提升 Scaling 能力，使其在总参数量 320B 与 GLM-4.5 相当的情况下，将激活参数量从 32B 降至 18B、层数从 92 减至 45，结合 30T Token 多模态预训练语料，用更少的计算资源实现超越 GLM-5.2 的性能。国产芯片推理引擎：针对国产芯片内存容量与带宽瓶颈，团队在 SGLang 基础上构建专用推理引擎，采用 EPD（Encode-Prefill-Decode）分离式架构将多模态编码、预填充和解码拆分为独立工作池，结合张量并行、ReplaySSM、W8A8 量化及混合缓存量化等技术，使端到端性能较基线提升 3 倍，达到与主流英伟达 GPU 相当的效率。微信关注回复“开源”，加入AI开源项目交流群如何使用GLM-5.3-FlashBigModel API：访问 BigModel 开放平台 https://docs.bigmodel.cn/cn/guide/models/vlm/glm-5.3-flash注册账号，获取 API Key 后按文档调用glm-5.3-flash接口。Z.ai API：访问 Z.ai  API 官网 https://docs.z.ai/guides/vlm/glm-5.3-flash文档中心，注册并获取 API Key 即可接入模型。Z.ai 在线聊天：访问Z.ai网页，登录后直接开始多模态对话。智谱清言 App：下载智谱清言 App并登录，在模型列表中选择 GLM-5.3-Flash 即可使用。ZCode：进入ZCode，启用模型后在代码编辑器中让 Agent 协同完成开发任务。AutoClaw：访问AutoClaw官网，上传文档让模型自动处理办公和专业文书任务。开源本地部署：从 HuggingFace 下载模型权重，按说明在本地服务器加载运行。GLM Coding Plan：前往 https://bigmodel.cn/glm-coding 申请限量体验卡，免费试用模型能力。GLM-5.3-Flash的核心优势极致性价比：与 Claude Opus 4.8 同级智能，定价仅为其 1/40，限时折扣内为 GLM-5.3 的 1/20。原生多模态：GLM-5 系列首个原生多模态模型，视觉能力原生集成，可自主判断并调用视觉反馈。高效架构：采用稀疏注意力与线性注意力混合架构，注意力计算量和 KV 缓存较 GLM-5.3 分别降低 3.01 倍和 4.44 倍。国产芯片部署：首次大规模使用国产芯片集群，端到端性能较基线提升 3 倍，效率对标主流英伟达 GPU。全面超越前代：用 320B 总参数 / 18B 激活参数，在多项基准测试中全面超越参数量高出一倍的 GLM-5.2。专业工作优化：针对金融、法律、Office 文档等专业场景深度优化，输出格式可直接交付。GLM-5.3-Flash的同类竞品对比维度GLM-5.3-FlashClaude Opus 4.8GLM-5.3总参数320B未公开约 355B激活参数18B未公开32B模型层数45 层未公开92 层AA 智能指数57 分57 分更高（约 60+）多模态能力✅ 原生视觉编码✅ 多模态✅ 多模态开源状态✅ 开源❌ 闭源❌ 闭源定价水平极低（Opus 4.8 的 1/40）高端定价中端定价（Flash 的 10~20 倍）注意力架构稀疏 + 线性混合标准 Transformer标准 Transformer长上下文支持1M200K+1M国产芯片部署✅ 已验证❌ 否❌ 否核心定位前沿智能普惠化顶级闭源模型上一代旗舰模型GLM-5.3-Flash的应用场景前端与游戏开发：根据视觉反馈自主迭代代码，生成可运行的网页、游戏和 3D 场景。金融专业工作：覆盖金融研究、报告生成、建模与分析全流程，提供可追溯的数据来源。法律实务处理：审查合同条款、批注修订，并起草符合实务规范的律师函与诉讼文书。Office 文档自动化：处理 PPTX、PDF、DOCX、XLSX 等格式，具备审美判断与排版优化能力。Agent 协同办公：通过 Browser Use Agent 和 Computer Use Agent 在浏览器、代码与 GUI 间自动协同。日常多模态对话：在 Z.ai 和智谱清言 App 中进行图文理解、创意生成与知识问答。# AI工具# AI项目和框架©版权声明本站文章版权归AI工具集所有，未经允许禁止任何形式的转载。上一篇Qwen3.8-Flash - 阿里通义推出的多模态 MoE 模型下一篇Hy-MT2-1.8B - 腾讯混元推出的端侧翻译大模型相关文章快问AI – AI驱动的覆盖多场景学习问题的中文学习助手AI小集2Open-o3 Video – 北大联合字节开源的视频推理模型AI小集2Perplexideez – 开源本地AI搜索助手，智能搜索信息来源追溯AI小集2袋鼠参谋 – 美团推出的商家AI智能决策应用AI小集3Penelope.ai – AI学术手稿检查工具，自动化检查确保手稿符合期刊要求AI小集2Fairies – 通用型AI Agent，强大的多任务执行能力AI小集6暂无评论再想想发表评论暂无评论…热门工具Loomy即梦AISekoAiPPT秘塔AI搜索妙呀堆友Agent美图设计室绘蛙AIupdream办公小浣熊秒哒最新收录AlphaVowXCellMetaDigChatArtLexcat造剧最新文章秒悟Meoo – 1分钟生成前后端完整的网站，真0门槛开发！9分钟前Rome – 开源的智能体操作系统，自然语言生成完整应用9分钟前Headlong – Laude Institute 开源的 AI Agent 微框架11分钟前Miya – 时域科技推出的 AI 音乐应用，画面可随歌词律动1天前Zing-0.5 – Loopit 开源的实时交互视频世界模型1天前2026 年 AI 生成图文店门头效果图 – 7 款工具深度横评2天前Hy4 preview – 腾讯混元开源的新一代旗舰大模型2天前Gemini Omni 1.1 Flash – 谷歌推出的 AI 视频生成模型2天前Twoo – 双人共用 AI 伙伴应用，能同时听懂双方对话2天前Prime Agent – Prime Intellect开源的长周期AI Agent运行框架2天前shuohao-skills – 开源AI短剧制作Skill集合，包含完整工作流2天前怎么用 MetaDig 做 AI 短视频，附教程实测案例3天前Gemini 3.5 Transcribe – 谷歌推出的最新语音转文本模型3天前Hy-MT2-1.8B – 腾讯混元推出的端侧翻译大模型3天前Qwen3.8-Flash – 阿里通义推出的多模态 MoE 模型3天前