🏆 Headline
Trillion Parameters, Weights Open: Mistral's Flagship Le Chonk Goes Open-Source
The open-source camp landed a big one today. Mistral announced it is releasing the full weights of Mistral Large 4 — the trillion-parameter (1T), natively multimodal flagship internally codenamed Le Chonk. It entered public preview earlier this month; today it ships open with three more weeks of reinforcement learning baked in, and blew past 1,000 points on Hacker News. The scorecard concentrates on two fronts: coding and defensive security. On Cognition's private FrontierCode 2.0 benchmark it leads all open-weight models at 19.7%. On Terminal-Bench 4.0 it climbed 13 points from the preview to 41% while cutting output tokens by 30%. On the Artificial Analysis Coding Agent Index it scores 48%, effectively tied with the leading open-weight model while using 47% fewer tokens per task. In Surge Code's blind human evaluations, its code ranked behind only Claude Opus 5.5 and ahead of the other open-weight models evaluated. The security numbers are written up as selling points too: a 96.6% refusal rate on malicious cyber requests — the highest Mistral measures among open models — and B3 agent-security attack resistance up from 93.3% in preview to 97.1%. Pricing and license came out together: $0.70 per million input tokens, $0.07 cached, $2.10 output. The license is a modified MIT — free to use, fine-tune and commercialize, with one condition: businesses operating a Model-as-a-Service offering need a separate commercial license. Trained from scratch on European infrastructure, the weights are on Hugging Face, also served via Microsoft Foundry, OpenRouter and NVIDIA-hosted endpoints.
Source: Mistral blog, TechCrunch | 2026-10-29
Reddit Announces Shift to Logged-In Experience: Search Traffic Dries Up, the Content Kingdom Starts Checking IDs
On its Q3 earnings call, Reddit announced a transition to a primarily logged-in experience. The data explains why: global daily active users were flat at 130.3 million, while US DAU fell from 53.2 million to 50.7 million — as AI-powered search rises, fewer visitors arrive from search engines. The higher the wall, the harder to scrape, and the more leverage in licensing talks with model providers.
Source: TechCrunch, Ars Technica | 2026-10-29
Thrive Buys ADT for $7.6B: OpenAI Supplies Both the Models and the Engineers
Thrive Holdings — the long-term-hold platform spun out of Josh Kushner's Thrive Capital — is acquiring security company ADT at $10 per share, roughly $7.6 billion in equity and $15.3 billion in total enterprise value. The structure is the story: OpenAI is both an investor in and a partner to Thrive Holdings, supplying models and forward-deployed engineers to its 75+ companies. ADT becomes the foundation of its smart home security business.
Source: TechCrunch | 2026-10-29
OpenAI Doubles Usage for All Pro Tiers: Effective Same Day, No Price Change
OpenAI announced doubled usage limits across its Pro 100 / 200 / 500 subscription tiers, effective October 29 with no price change, plus API credits for eligible subscribers — confirmed on X by the company's Codex lead. Last week the conversation was usage anxiety; this week the answer is simply more.
Source: OpenAI (X), multiple outlets | 2026-10-29
Google Winds Down Jules: the Agent Coding Tool Exits on December 31
Google announced in its official changelog that the Jules coding agent will wind down on December 31, 2026, with users having until November 11 to wrap up sessions; the official framing is "unifying our agentic coding efforts." The asynchronous coding agent that grew out of Google Labs ends its run a little over two years in.
Source: Jules official changelog | 2026-10-29
Deepgram Acquires Daily: Pipecat Joins, and Voice AI Assembles a Full Stack
Voice AI company Deepgram announced it is acquiring Daily — the maker and maintainer of Pipecat, one of the most popular open-source voice agent frameworks. Terms undisclosed; Pipecat stays open source. Pipecat orchestrates voice and video streams along with speech-to-text, LLMs and text-to-speech while keeping latency down; buying it lets Deepgram combine "hears well, answers fast" with "connects to anything" in one stack.
Source: TechCrunch, Deepgram blog | 2026-10-29
Wabi Raises $55M at $1.25B: Replika's Founder Bets on the Personal Agent
Wabi, founded by Eugenia Kuyda (creator of Replika), raised $55 million at a $1.25 billion valuation. Last month the company pivoted from "build apps with prompts" to an AI messenger with a personal agent: Wabi 2.0 generates the small tools you need mid-conversation — meal planners, restaurant finders, fitness plans, trivia games.
Source: TechCrunch | 2026-10-29
Rivian Licenses Its Voice Assistant to Volkswagen: the $5.8B JV Gains Another Layer
Rivian confirmed in its Q3 investor presentation that its $5.8 billion joint venture with Volkswagen Group is expanding: elements of Rivian's AI voice assistant technology will be licensed to VW. Fees and launch vehicles weren't disclosed, but the likely first stop is VW's budget EV, the ID EVERY1 — already the first VW to use Rivian's electrical architecture.
Source: TechCrunch | 2026-10-29
🏆 今日头条
万亿参数权重全开放：Mistral 旗舰 Le Chonk 转身开源
开源阵营今天接了个大的。Mistral 宣布公开旗舰模型 Mistral Large 4 的全部权重——这个内部代号 Le Chonk 的万亿参数（1T）原生多模态模型，本月初刚以公开预览的身份亮相，如今带着三周强化学习的成果直接开源，在 Hacker News 上冲破 1000 分。 成绩单集中在编码和防御性安全两条线上。Cognition 的 FrontierCode 2.0 私有基准上，它以 19.7% 领跑全部开放权重模型；Terminal-Bench 4.0 从预览版的 28% 涨到 41%，输出 token 还少了三成；在 Artificial Analysis 的编码智能体指数上拿到 48%，与最强的开放权重模型基本打平，但每个任务平均少用 47% 的 token。人工盲测里，它的代码被评价者排在 Claude Opus 5.5 之后、其余开放权重模型之前。安全一侧的数据同样被写进了卖点：恶意网络请求的拒绝率达 96.6%，是官方测过的开放模型里最高的；B3 智能体安全基准对越界尝试的抵抗率从预览版 93.3% 升到 97.1%。 价格与许可证一起公布：每百万 token 输入 0.70 美元、缓存输入 0.07 美元、输出 2.10 美元；许可证是改版 MIT——自由使用、微调、商用，唯一条件是转售模型服务（Model-as-a-Service）的商家需另行取得商业授权。模型在欧洲从头训练，权重已上架 Hugging Face，也可经 Microsoft Foundry、OpenRouter 或 NVIDIA 托管端点使用。 > 💬 开源已经不是「追平闭源」的证明动作，而是正经的商业模式：权重送出去，换回部署权和微调生态，再用一条「转售需授权」的条款把平台型买家请回谈判桌。编码加安全的组合拳，瞄的是企业预算里最舍得花钱的两类智能体负载。
来源：Mistral 官方博客、TechCrunch | 2026-10-29
Reddit 宣布转向登录制：搜索流量见底，内容王国开始收门
Q3 财报电话会上，Reddit 宣布转向以登录体验为主的模式。理由写在数据里：全球日活环比持平在 1.303 亿，美国日活从 5320 万掉到 5070 万——AI 搜索兴起后，从搜索引擎漏进来的访客越来越少。按公司的说法，墙砌高一点，内容更难被抓取，与模型厂商谈数据授权时的筹码就多一分。 > 💬 免费互联网的第一块多米诺：当增长引擎从「搜索带来的流量」换成「授权费」，登录墙就是新时代的付费墙——这次收的不是钱，是你的身份。
来源：TechCrunch、Ars Technica | 2026-10-29
Thrive 76 亿美元买下 ADT：OpenAI 既出模型又出工程师
Thrive Holdings——Josh Kushner 的 Thrive Capital 分拆出的长期持有型平台——宣布收购安防公司 ADT：每股 10 美元，股权对价约 76 亿美元，企业总价值 153 亿美元。特别的是交易结构：OpenAI 既是 Thrive Holdings 的投资方也是其合伙人，向旗下 75 家以上的公司输出模型与前驻工程师。ADT 将成为其智能家居安防业务的底座。 > 💬 「资本 + 模型厂」合体收购传统硬件公司，是 AI 落地最重的形态：资本买壳，模型进场，工程师驻场。
来源：TechCrunch | 2026-10-29
OpenAI 给 Pro 用户全线翻倍用量：当晚生效，价格不变
OpenAI 宣布 Pro 100 / 200 / 500 三档订阅的用量上限全部翻倍，10 月 29 日当天生效、价格不变，部分订阅还会获得 API 额度，Codex 负责人在 X 上确认了消息。上周大家还在讨论限额焦虑，这周直接给量。 > 💬 限额是成本函数：模型每便宜一点，用户就多得一截。翻倍不涨价，比任何发布会都诚实地标出了推理降本的斜率。
来源：OpenAI（X）、多家媒体 | 2026-10-29
Google 关停 Jules：智能体编码工具 12 月 31 日谢幕
Google 在官方更新日志中宣布，智能体编码工具 Jules 将于 2026 年 12 月 31 日关停，用户有到 11 月 11 日的时间收尾未完成的会话；官方口径是「统一我们的智能体编码工作」。这款从 Google Labs 走出的异步编码智能体，生命定格在两年出头。 > 💬 「统一」是体面的说法，其实就是收敛赌注。编码智能体的牌桌已经容不下试错位——活下来的选手开始清场。
来源：Jules 官方更新日志 | 2026-10-29
Deepgram 收购 Daily：Pipecat 并入，语音 AI 凑齐全栈
语音 AI 公司 Deepgram 宣布收购 Daily——广受欢迎的开源语音智能体框架 Pipecat 的开发与维护方，金额未披露，Pipecat 承诺继续开源。Pipecat 负责串起语音视频流与语音识别、大模型、语音合成的整条管线并压低延迟；买下它，Deepgram 把「听得清、答得快」和「连得上」拼成了一套完整栈。 > 💬 语音智能体赛道的整合开始了：框架层是入口，谁握住开源默认项，谁就定义别人的接线方式。
来源：TechCrunch、Deepgram 官方博客 | 2026-10-29
Wabi 新融 5500 万美元：Replika 创始人的「个人智能体」估值 12.5 亿
Eugenia Kuyda（Replika 创始人）创办的 Wabi 宣布以 12.5 亿美元估值完成 5500 万美元新融资。公司上月刚把定位从「用提示词生成应用」转向「AI 信使 + 个人智能体」：Wabi 2.0 会在对话里即时生成你要的小工具——排餐单、找餐厅、定健身计划、出题小游戏。 > 💬 「即用即弃的小应用」是个聪明的切口：应用不再被下载，而是被生成。估值能否撑住，就看生成的东西第二次还敢不敢用。
来源：TechCrunch | 2026-10-29
Rivian 把车载语音助手授权给大众：58 亿合资再加一层
Rivian 在 Q3 投资者材料中确认，与大众集团 58 亿美元的合资范围扩大——其车载 AI 语音助手技术将授权给大众使用。授权费用与落地车型未披露，首发大概率是大众的平价电动车 ID EVERY1：那也是第一款采用 Rivian 电子电气架构的大众车型。 > 💬 软件定义汽车走到下半场：车企间的关系正在变成「我造车，你租我的助手」。
来源：TechCrunch | 2026-10-29