Qwen Teases Its Next-Generation Flagship Model with a Poster
模型发布
Qwen’s official account released a teaser poster centered on the fish Kun from the Zhuangzi, leading the community to read it as a sign of an approaching next-generation flagship model, while its parameters, open-source status, and launch date remain undisclosed.
Alibaba’s official Qwen account published a poster displaying “北冥有鱼 其名为_” and used The becoming begins as its only caption. Qwen Developers later explained that the Chinese text comes from the ancient Chinese classic Zhuangzi and refers to Kun, a fish living in the northern sea. Its pronunciation matches Qwen, while its body is said to be so large that nobody knows how many thousands of li it extends. Community users widely interpreted the post as a teaser for the next-generation flagship model, but the official accounts have not stated its parameters, whether it will be open source, or when it will launch.
Alibaba’s MNN team has open-sourced MNN-Pocket-0.3B for smartphones. The model has about 300 million parameters, supports Chinese and English instructions and up to 32K tokens of context, while its INT4 deployment package is 173 MB.
Alibaba’s MNN team has released and open-sourced MNN-Pocket-0.3B, a model designed for the limited computing capacity and memory available in smartphone applications. The dense language model has about 300 million parameters and supports instruction-based interaction in Chinese and English. It supports context lengths of up to 32K tokens and is distributed under the Apache 2.0 license. The INT4-quantized MNN deployment version has a size of 173 MB.
According to the team, when MNN-Pocket-0.3B runs offline on an iPhone 18 Pro through the MNN engine, its average decoding speed is about 350 tokens/s.
Qoder CN Launches a Limited-Time, Limited-Quantity Free Offer for the Anonymous Model Qoduck
开发生态
Starting November 7, 2026, Qoder CN will give all new and existing individual users a limited daily free quota for the anonymous Qoduck model. Usage within the quota will carry a 0.0× coefficient instead of 0.5× and consume no Credits; the end date is undecided.
Qoder CN will start a time-limited free campaign for Qoduck on November 7, 2026. The end date has not been decided, and Qoder CN will announce it in advance through official channels. The model’s identity remains undisclosed, and individual users can use it simply by selecting the model without claiming a quota. Trial users, free accounts, all individual subscription tiers, membership-card users, and accounts with a Credits balance of 0 are eligible. Team, Enterprise Standard, and Enterprise Dedicated plans are excluded, and Qoduck will not appear in those users’ model selectors.
Each account receives a fixed number of free uses per day, with paid subscribers receiving more than free users; the quota returns the following day after being exhausted. Qoduck calls within the quota consume no Credits, and the model is also charged at 0.0× when the thinking setting is set to “极高,” or “最大” in another FAQ section. After the quota is used up, users may continue calling the model by spending Credits at its original 0.5× coefficient. Qoder CN may change the quota or campaign period based on resource conditions, and peak periods may bring slower responses or queues. Qoder CN IDE, the Qoder CN JetBrains plugin, and Qoder CN CLI are not supported.
Google Antigravity Resets Quotas and Launches Voice Mode
开发生态
Google Antigravity has launched a Gemini Live-powered voice mode in the Antigravity App and CLI. Users can assign tasks by voice, steer the Agent in real time, and handle some permission requests without returning to the keyboard.
Google Antigravity has made voice mode, powered by Gemini Live, available in the Antigravity App and Antigravity CLI. According to the official explanation, users can give the Agent tasks by voice, add instructions while work is running, ask for status updates, stop tasks, and answer permission requests. The Agent responds aloud in real time while continuing to read code, edit files, run commands, and show results in the same conversation. The context of the existing conversation is retained.
Voice approvals remain subject to the existing permission rules. Destructive actions, including deleting files, running commands with elevated privileges, and force-pushing to a Git remote, cannot be approved by voice alone; users must click Allow on screen. Voice mode offers nine voices and speech speeds from 0.75x to 2x. It starts from the message box in the App, or with /voice live or F6 in the CLI. Task results, edits, and a transcript of both sides of the conversation remain in the conversation.
Google Redesigns Antigravity and Introduces Remote Task Support
开发生态
Google has redesigned the Antigravity app and opened remote agents to all paid personal subscription plans. Tasks can run in independent cloud virtual machines, so users do not need to keep their computers turned on.
Google has completed a redesign of the Antigravity app and introduced remote agents. According to the company, the feature is available across all paid personal subscription plans. Users can start tasks from anywhere, while each task runs in an independent cloud virtual machine, removing the need to keep a local computer powered on during execution.
Within a single project, users can also connect a GitHub repository and a Google Drive folder at the same time, then read and write code and documents in one conversation. Work on a project can continue between the web and the app.
TRAE Accounts to Be Upgraded to 豆包飞书 Accounts with Data Benefits Preserved
开发生态
TRAE said its accounts will be automatically upgraded to Doubao Feishu accounts on November 18, 2026, while existing data rights are retained. The upgraded accounts can also be used with other products and services that support the same credentials.
According to TRAE, the account-system integration will take effect on November 18, 2026, when existing TRAE accounts will be migrated to Doubao Feishu accounts. Related data and user rights will be retained under the terms of the announcement. After the migration, the upgraded accounts can be used with other products and services that support Doubao Feishu account credentials. The Doubao Feishu account service will be provided by Beijing Chuntian Zhiyun Technology Co., Ltd. and/or its affiliates.
The official explanation states that the detailed integration plan and rules are set out in the full announcement. Users with questions or objections can contact TRAE or the Doubao Feishu account service provider through feedback@mail.trae.cn or feedback@mail.doubao.com.
Many Users Report Rising DeepSeek Usage Costs in WorkBuddy
开发生态
As of November 7, 2026, many users reported higher point consumption when the China version of WorkBuddy calls DeepSeek-V4.1-Flash, but no announcement or confirmation of a pricing change has been issued.
As of November 7, 2026, multiple users of the China version of WorkBuddy reported that calling the DeepSeek-V4.1-Flash model consumed more points than before. Users observed that the point price displayed in the software interface had not changed. Users who contacted relevant staff were told that no notice of a price increase had been received. At this stage, information about a possible price change for the model is still based on user observations, and the platform has not issued a related announcement.
OpenDesign Launches DeepSeek V4.1 Flash Eco with $220 in Monthly Credits
开发生态
OpenDesign has launched DeepSeek V4.1 Flash Eco for batch and non-real-time work. OpenDesign Go includes $220 in monthly Eco usage, while the standard version remains available through the official DeepSeek API with token-based billing.
OpenDesign is making DeepSeek V4.1 Flash Eco available for batch processing and non-real-time tasks. According to the company, Eco uses the same weights as DeepSeek V4.1 Flash and retains the same agent, programming, reasoning, and multimodal capabilities, with support for a 1-million-token context. Its response priority is lower, however.
OpenDesign Go costs $8 for the first month and $10 per month thereafter, including $220 of Eco usage each month. Eco is also available through Go Plus and the API. The standard version continues to be offered by OpenDesign through the official DeepSeek API and is billed by token usage, with a higher response priority than Eco.
Command Code has launched DeepSeek V4.1 Flash 1017 Ultrafast, with a 1-million-token context window and throughput exceeding 1,000 tokens per second. It is available through GOAT and higher-tier plans and via API.
Command Code has launched DeepSeek V4.1 Flash 1017 Ultrafast. The model offers a 1-million-token context window and throughput of more than 1,000 tokens per second, and can be accessed through GOAT and higher-tier subscription plans as well as the API.
GOAT costs $10 per month and includes $20 in credits. The company’s model pricing page lists DeepSeek V4.1 Flash 1017 Ultrafast at $2.40 per 1 million input tokens, $9.60 per 1 million output tokens, and $0.05 per 1 million cached-read tokens.
Anthropic has added Claude Code session features to the latest App Store version of Claude Duet, allowing users to check tasks running locally on a computer or in the cloud, approve changes, and send updates from a phone. The rollout is gradual.
Anthropic enabled Claude Code session features in the latest App Store version of Claude Duet, with the update now rolling out gradually. Users can use a phone to view tasks in progress, approve code or other changes, and send updates to a session. The feature supports sessions running locally on a computer as well as sessions running in the cloud.
Claude Code Optimizes Cloud Session Environment Configuration
开发生态
Anthropic has optimized Claude Code’s cloud-session environment setup so Pro, Max, and Ultra users can have required dependencies, network access, and API keys identified while preserving the environment for later sessions.
Anthropic has changed the environment configuration for Claude Code cloud sessions. According to the company, when Pro, Max, and Ultra users run a code repository, Claude identifies the required dependencies, network access, and API keys, then saves the configured environment for use in subsequent sessions. API keys are stored securely, and Claude cannot view them.
OpenRouter Adds Parallel Decisions with a 7-Day Free Trial
开发生态
OpenRouter has listed Parallel’s Decisions Fast (free), released on November 6, 2026, with support for up to 100 questions and a 32,000-token context window at zero cost.
OpenRouter lists Parallel’s Decisions Fast (free) as an available model served by one provider. It reads caller-provided text or JSON as state, researches the live web, and returns structured results such as the probability of yes for yes/no questions, choice with probabilities for each option, and score values on an ordered rubric. It can refuse individual questions and attach citations to sources used for the request. The model is intended for fact checks, entity screening, and classification that depend on current information; a request can contain up to 100 questions about the same state, with a limit of 32,000 tokens per question including the shared state. The page states that prompt and completion tokens cost zero, while free endpoints are rate limited. The model is not currently available through EU in-region routing. Calls must use the OpenRouter Decisions API, and chat completions SDKs for the OpenAI-compatible chat endpoint do not work with it.
Cursor launched real-time voice chat on November 6, 2026. Users can speak directly with an Agent in the Agents Window, IDE, web, and iOS while keeping the Agent working on coding tasks.
Cursor enabled real-time voice chat on November 6, 2026. According to the company, users can describe code changes aloud, ask what an Agent is doing, and continue the conversation while a task runs. The feature is available in the Agents Window, the IDE, on the web, and in Cursor for iOS. When an Agent is idle and the chat input is empty, users can start a call by clicking the waveform button labeled Start voice chat, which appears where Send normally sits. On iOS, saying “Hey Siri, call Cursor” can start a call with a new Agent, while a Live Activity on the Lock Screen provides mute and hang-up controls.
The voice Agent listens to the request and hands coding work to the Agent in the chat. During the run, users can ask about changes in a diff, check a run, search earlier messages and calls with that Agent, or switch between workers in a Project that contains them. According to the company, the coding Agent gives short spoken progress updates and a brief response when the work is complete. Voice chat runs on Grok Voice, with voices including Ara, Eve, Leo, Rex, and Sal available in Voice settings. With Privacy Mode enabled, voice chat runs with zero data retention. Voice chat is separate from dictation, previously called Voice Mode, which still converts speech into text in the chat input; Cmd+Shift+Space continues to toggle dictation in the IDE.
Amp Launches Spaces, Letting One Main Agent Coordinate Multiple Threads
开发生态
Amp has launched Spaces, where a primary Agent creates, assigns, and coordinates multiple work threads so teams can handle development and operations in one shared space without organizing conversations manually.
With the launch of Spaces, Amp introduced a workflow in which one Agent manages work carried out by multiple Agents. Users specify a project and goal in a space, after which the primary Agent creates threads for the relevant tasks and coordinates their execution. Team members can communicate through comments in the space and tag Amp when they are ready for action. According to Amp, its team used a space called “Spaces v1” to collect feedback, bugs, and new ideas before having the Agent delegate them into threads. Lewis used one space to roll out service accounts, monitor X, Slack, and bug reports, fix incoming issues, and send emails and refunds to affected users. In another space, a security-review task split the codebase into 16 areas, reviewed each area in its own thread, checked every finding in another thread, and retained a summary artifact. A space starts from the “+” in the Spaces section of the sidebar by setting a title and project, then telling the Agent what to work on.
Doubao Work Personal removed its previous 5-hour usage limit on November 6, allowing users to schedule usage within their weekly quota; paid users can still receive quota-reset cards periodically.
Doubao Work Personal changed its quota rules on November 6 and removed the 5-hour per-session usage cap. Previously, users had to wait for a short-cycle quota reset after reaching the 5-hour limit, even when part of their weekly quota remained. Under the new arrangement, usage time can be allocated according to the weekly quota and each user’s work schedule. The weekly quota and other usage rules continue to depend on the applicable plan, while paid users may also receive quota-reset cards from time to time. Users can check their plan, remaining quota, and usage rules under “Settings — Subscription and Quota Management.”
Notion Launches a ChatGPT Plugin for Editing Pages Alongside Conversations
产品应用
Notion has introduced a ChatGPT plugin panel that lets users open or create Notion pages beside a conversation, edit them directly or ask ChatGPT to make changes, with support for the web and desktop app.
Notion has introduced a panel for opening, creating, and editing pages beside a ChatGPT conversation. Users can type directly on a page or ask ChatGPT to edit it, create a database, or add a comment. The panel uses the same connection as Notion MCP, and users must install the Notion plugin in ChatGPT, sign in to Notion, and choose a workspace. Existing Notion permissions still apply: users can open only pages they can access and edit only pages where they have edit access.
The panel connects to one workspace at a time. Unsaved changes on the device are not retained if the connection drops, and editing is unavailable offline. Member management, billing, connection settings, sharing permissions, and link previews that require a connected account must still be handled in Notion. The panel is enabled by default, and on the Enterprise plan a workspace owner can disable it for everyone through Settings → Notion MCP → Manage workspace clients. After it is disabled, members can still use Notion through ChatGPT chat if ChatGPT is allowed to connect.
Meta Announces New Dedicated Models for Muse and Meta AI
产品应用
Meta has begun rolling out specialized models trained with Muse Wonder to Muse and Meta AI users, adding Active Memory and Active Focus. In Muse tests, forgetting critical information fell 2.4x, while completion of everyday tasks involving unexpected issues improved by 67%.
Meta has begun providing Muse and Meta AI with specialized models trained using Muse Wonder. According to the company, Muse Wonder targets complex Agent tasks in work and everyday life and was also used to improve the models behind both products. The current Muse model scores 47% on the Personal Superintelligence Index, compared with 38% for the previous model.
The new models include Active Memory and Active Focus. Active Memory identifies whether information is temporary or likely to remain useful, then retrieves relevant details in later conversations. According to Meta, Muse showed a 2.4x reduction in forgetting critical information needed to achieve goals, while its share of tasks correctly using critical context on the Active Memory eval reached 90%, up from 76%. Active Focus allows the model to keep pursuing a goal when unexpected obstacles arise, and Muse showed a 67% improvement in completing such everyday tasks. Meta also said the models can handle multitasking, decide when to advance goals independently or collaborate live, and respond to changes and clarifications during task execution. In Meta AI, the improvements support breaking complex problems into steps, asking follow-up questions, and adapting when the conversation changes direction.
Meta Launches Spaces for Muse to Create Custom Experiences Through Conversation
产品应用
Meta has launched Spaces in Muse, allowing users to describe what they need in conversation and have the system create custom experiences such as planners, trackers, and games. The feature is available in every region where Muse is offered.
Meta has added conversational creation and editing to Spaces in Muse, and the feature is now available in all regions where Muse is offered. Users can describe the experience they want and have Muse build a custom space such as a planner, tracker, or game. After creation, a space can be placed in the sidebar or main tab, while further conversations with Muse can change its appearance and functions. After commonly used apps are connected, Spaces can update information such as event dates and price changes in the background.
Anthropic Introduces Claude Dreams, Which Organizes Memories in the Background, Now in Research Preview
技术与洞察
Anthropic introduced Claude Dreams in a research preview, allowing Claude to organize session logs between interactions, validate experience, and save it as memory for later sessions. In CL-Bench 1.0, Claude Opus 5 and some later models formed learning loops using that memory.
In its research preview, Anthropic defines Claude’s Agent memory as knowledge gained from user interactions and assigned tasks that is saved as context Claude can read in a later session. It excludes changes to model weights, in-context learning within a single conversation, and compaction, which replaces older turns with summaries. Claude can tag salient events during a session or write them to a memory store, then use background processing resembling “dreaming” to review logs, correct mistakes, and produce rules and notes.\n\nIn the CL-Bench 1.0 database-question task, each question began in a fresh session, while the models could access a memory directory preserved across sessions. Claude Sonnet 4.6 generated theories about why answers were wrong; Claude Opus 4.7 consolidated notes into a reference document and separated checked from unconfirmed facts; Claude Sonnet 5 recorded confirmed answers and corrections; and Claude Opus 5, Claude Opus 5.5, and Claude Sonnet 5.5 stored validated rules to form learning loops for later questions. Anthropic also evaluated memory selection in Pokémon Red. At about 10,000 steps, Claude Sonnet 3.7 had written 413 files across 225 folders and earned one badge, while Claude Fable 5 had written nine files in two folders plus a top-level global.md.
Anthropic Report: AI Could Double the Output of Some Scientific Research
技术与洞察
An Anthropic report analyzed researchers’ use of Claude, Claude Code, and Claude Science, stating that current capabilities could at most double research output in some fields; if AI fully automated non-physical work, output in fields such as computer science and statistics could rise by up to 15 times.
An Anthropic research report analyzed how researchers use Claude, Claude Code, and Claude Science and estimated the potential effect of AI on the pace of scientific research. According to the report, formal sciences such as computer science, mathematics, and statistics have higher usage relative to their number of researchers, while health and laboratory-based physical sciences have lower usage. By task, researchers use Claude more for data analysis and writing, and less for data gathering or generating ideas and hypotheses. The report states that current AI capabilities could increase research output in some fields to as much as double, mainly through task-level augmentation. In a hypothetical scenario where AI fully automates all non-physical work, output in highly cognitive fields such as computer science and statistics could increase by up to 15 times.
Physical Superintelligence Claims to Have Proved a Conjecture in Quantum Computing
技术与洞察
Physical Superintelligence released a proof, paper, and Lean 4 formalization of the Aaronson–Ambainis conjecture on GitHub, including an explicit influence bound. AI models produced the proof code, which was rechecked by two independent Lean kernel implementations.
Physical Superintelligence has published a proof of the Aaronson–Ambainis conjecture, its paper, and a Lean 4 formalization in a GitHub repository. The conjecture concerns bounded low-degree polynomials on the discrete cube: if the polynomial’s variance is at least ε, at least one variable has an influence lower bound. On {−1,1}^n, Theorem 1.1 in the repository’s paper gives the bound (√15 / (500 π^4)) · Var(f)^{5/2} / d^6. The conditions are n ≥ 1, function values in [−1,1], and degree at most d ≥ 1.
According to the repository, all four theorems have been proved in Lean without placeholders or added axioms, and rechecked by two independent implementations of the Lean kernel. The Lean version is pinned to Lean 4.33.0 and the dependency version to Mathlib v4.33.0. AI models produced the complete development, the first draft of the paper, and the formalization in a research harness developed by David Turturean at PSI; Turturean reviewed the paper and is responsible for its correctness. The code, scripts, and documentation are released under the Apache License 2.0.
OpenAI and Ligo Launch a Benchmark for Biological Experiment Analysis
技术与洞察
OpenAI and biotech company Ligo launched BioBasicsBench to evaluate AI agents on 100 real wet-lab experiment tasks. In the xhigh setting, GPT-6 Astra scored 66.7% and GPT-5.6 Sol scored 43.8%, while OpenAI said the results do not mean the models can make reliable laboratory judgments.
OpenAI launched BioBasicsBench with biotech company Ligo to test whether AI agents can handle routine analysis work in wet laboratories. The benchmark contains 100 real experimental tasks and examines whether agents can interpret raw data, identify errors, and determine whether the available evidence is sufficient to support a conclusion.
With the xhigh setting, which allows more reasoning for each task, GPT-6 Astra achieved an average usable-score rate of 66.7%, compared with 43.8% for GPT-5.6 Sol. OpenAI said that a higher benchmark score does not mean a model can already make reliable judgments in a laboratory.
Artificial Analysis Launches the VIALS-AA Leaderboard for Evaluating Scientific Image Reasoning
技术与洞察
Artificial Analysis launched the VIALS-AA leaderboard to evaluate how models interpret life-science research images using Handshake AI’s VIALS benchmark. Gemini 4 Carbon led with 64%, followed by GPT-6 Astra at 63% and Claude Opus 5.5 at 61%.
The VIALS-AA leaderboard launched by Artificial Analysis uses Handshake AI’s VIALS benchmark and covers 161 tasks across eight fields. Each task is tested three times, and the final score represents the share of tasks answered correctly in all three trials. Artificial Analysis said Gemini 4 Carbon scored 64%, while GPT-6 Astra and Claude Opus 5.5 scored 63% and 61%, respectively.
GPT-6.1 Sol scored 59% and Terra scored 60%; their average cost per task was approximately $0.056 and $0.073, respectively. Kimi K3.1 Pro scored 45%, placing first among open-weight models. The leaderboard evaluates models’ ability to interpret images from life-science research.
