DeepSeek reportedly secures at least 80 billion in new funding; Microsoft unveils a Jev‑style decision model with blazing speed.
DeepSeek reportedly raises at least 80 billion; Microsoft releases Jev-like decision model, speed…
Summary
Compiled | Edited by Lin Qibei | Cai Fangfang NVIDIA negotiating acquisition of Reflection AI: the latter just released Beam, targeting Chinese open-source models; Doubao responds to integrating utility payment features: building for life scenarios
On October 11, according to the Financial Times citing sources familiar with the matter, Nvidia is in preliminary talks with U.S. AI startup Reflection AI to discuss the possibility of an acquisition or expanding its existing investment. Potential deal structures include a full acquisition, increasing equity investment, and a talent acquisition approach by hiring the core team and obtaining technology licensing. The negotiations are still in early stages, and no agreement has been reached yet.
Reflection AI was founded in 2024 by former Google DeepMind researchers Misha Laskin and Ioannis Antonoglou, focusing on developing open-weight AI models. The company was previously valued at $25 billion, and Nvidia has invested about $800 million to date, making it one of its major investors.
Just before the acquisition news broke, Reflection AI released its first open-weight model, Beam, on October 5. The model uses a mixture‑of‑experts (MoE) architecture, has a total of 501 billion parameters and 23 billion active parameters, and is primarily aimed at code generation, complex reasoning, and AI Agent tasks. Reflection AI’s published test results show that Beam’s performance on certain reasoning benchmarks is close to Zhipu GLM‑5.2, with an estimated inference compute cost roughly one‑quarter to one‑third of the latter. Reflection AI positions Beam as a competitive product against Chinese open models such as DeepSeek, GLM, and Qwen.
On October 9, regarding reports that “Doubao can pay water, electricity, and gas bills,” a near‑Doubao source told the media that the company is currently building the capability to handle everyday payment scenarios, and the related functions are still being explored and refined.
Earlier, some users discovered a life‑services entry in the Doubao app and tried to complete certain daily tasks through conversational interactions. The features cover everyday service scenarios such as utility payments and travel, allowing users to describe their needs in natural language and receive corresponding AI‑driven support.
Previously, Doubao added services such as flight booking, train ticket purchase, ride‑hailing, and route navigation, and launched a dedicated “Travel with Doubao” entry. So far, Doubao has not announced the specific launch date, coverage scope, or partner information for the utility‑payment feature.
On October 7 (US time), OpenAI launched GPT‑6 and Intelligent UI, expanding GPT‑6’s user coverage. The most noticeable change in this update is the response format: previously, ChatGPT mainly generated text. Intelligent UI changes this model, allowing ChatGPT to combine text, images, charts, buttons, and forms into directly operable interfaces based on the question. For example, product comparisons can be displayed side by side, knowledge explanations can generate interactive diagrams, and calculation or planning tasks can provide tools directly. OpenAI says the system will choose the presentation mode according to the task, and questions suited to textual explanations can still be answered with plain text.
According to the official release notes, GPT-6 also changes the previous “think first, then answer” interaction model. Now it can start responding while still thinking or invoking tools, and gradually supplement the result, so users don’t have to wait for the entire process to finish before reading.
It should be noted that some models in the GPT-6 series have already been released; this week’s new developments are a version update for everyday conversations, the launch of Intelligent UI, and an expanded coverage. According to OpenAI’s announcement, GPT-6 with Intelligent UI will be gradually rolled out worldwide to Plus, Pro, Business, and Enterprise users starting on the day of release. The underlying model differs by subscription tier. Plus, Pro, Business, and Enterprise users will be supported by GPT-6 Sol starting October 7, while Free and Go users will receive GPT-6 Luna and the corresponding interactive experience starting October 8. Availability for enterprise users depends on administrator settings.
Recently, multiple media outlets reported that DeepSeek is nearing the close of a new financing round, raising at least ¥80 billion (about $12 billion). According to the signed term sheet, the final amount could approach ¥100 billion, far exceeding the original target of roughly ¥50 billion. CATL and Tencent Holdings are the lead investors in this round.
According to reports, DeepSeek launched this round of financing in July this year, targeting a valuation of about 500 billion RMB. Sources say that with the release of the company’s latest AI model, investor demand exceeded the original target. This round of financing is also another large‑scale fundraising for DeepSeek after the first external round completed in June this year. Previously, DeepSeek’s first round raised over 50 billion RMB, with investors including Tencent, CATL and other firms.
AI company TypeSafe announced the completion of an $870 million Series A round, bringing its valuation to $7.5 billion. The round was led by Andreessen Horowitz (a16z), with participation from Sequoia Capital, early investor DCVC, and several angel investors. a16z partner Martin Casado will join TypeSafe’s board of directors.
TypeSafe is the development company behind the AI model Jev, founded in 2024. Its team includes researchers who have previously worked at OpenAI, Meta, and Google. The recently released Jev has gone viral across the internet; it is positioned as a decision model for software systems. Unlike traditional large language models that primarily generate natural language text, Jev does not output textual content but directly produces structured decision results, including binary judgments, option selections, and scoring formats. The company states that Jev has been adopted by some enterprises and claims that roughly one‑third of Fortune 500 companies are currently using the model.
This financing round comes just a few weeks after Jev’s launch. TypeSafe says the new funds will be used to further refine the Jev product, develop more models geared toward machine invocation, and build infrastructure to support enterprise deployments.
On October 8, the parent company of AI agent firm Manus, Butterfly Effect, announced that it had recently completed a new financing round of over $500 million. The round was led by Boyu Investment and IDG Capital, with existing shareholders Tencent, Sequoia China (HSG), and ZhenFund continuing to invest. This is the first financing announcement since Manus terminated its acquisition deal with Meta and resumed independent operations. On September 1 of this year, Manus officially announced the resumption of independent operations.
In September this year, Manus released Manus 2.0 and launched the standalone personal agent application Cue. Manus 2.0 updates the existing agent architecture, while Cue is aimed at everyday personal tasks, supporting the configuration of a dedicated email, phone number, digital wallet, and computing environment for the AI Agent, enabling it to perform tasks across services and complete payments and other actions within the permissions set by the user. In addition, Manus announced that it is forming a domestic product team to advance cooperation with Chinese large‑model vendors and ecosystem partners.
On October 8, three former OpenAI safety researchers—Tomek Korbak, Jasmine Wang, and Mikita Balesni—co‑authored an open letter questioning the company’s decision to fire them. The letter, addressed to OpenAI’s Safety and Security Committee, Safety Advisory Group, and Mission Advisory Committee, warned that the dismissals could create a “chilling effect” on internal safety discussions, causing employees to hesitate to raise safety concerns publicly or to collaborate with external assessment bodies.
The three disclosed their individual circumstances in the open letter and personal statements. Balesni said, “The reason we were terminated is that we placed AI safety above OpenAI’s short‑term corporate interests.” Korbak noted that he had repeatedly warned that OpenAI was gradually losing the ability to monitor the reasoning processes of AI agents and expressed concern that the company was cutting back its collaboration with the external safety assessment organization METR. Wang stated that the company’s stated reason for her dismissal involved her accessing a senior executive’s email. All three argue that their work was related to maintaining AI safety and call on OpenAI to preserve the monitorability of frontier models, support independent safety assessments, and protect employees’ right to voice dissent.
On October 9, OpenAI responded that, after an internal investigation, it determined that three individuals violated the sensitive information access and handling policy, constituting a serious breach of trust. OpenAI emphasized that the dismissal decision is unrelated to the employees raising safety concerns or expressing dissent, and the issues uncovered by the investigation go beyond what the three disclosed in the open letter. The company also stated that it is advancing a cooperation agreement with a third‑party security assessment firm and will continue research on model observability.
On October 8, Anthropic announced its first major revision to Claude’s Usage Policy in more than a year, with the new rules taking effect on November 12, 2026. Among the updates, Anthropic added a clause prohibiting users from engaging in “sustained and needless abusive or cruel behavior” toward Claude.
Anthropic clarified that this provision targets extreme cases of repeated, purposeless mistreatment of the model and does not restrict users from expressing legitimate dissatisfaction, critiquing model performance, testing its capabilities, or exploring sensitive topics in creative, research, or other contexts.
Under the new policy, Anthropic still permits model evaluation, safety testing, and other research activities. For the few dialogues that involve ongoing abusive behavior, Claude could already end the conversation proactively; this update formally incorporates such behavior into the usage policy.
In addition to adding new clauses related to “AI abuse,” this policy update also covers political activities, elections, weapons development, surveillance, and applications in high‑risk domains. Anthropic says that as Claude is used for longer, more autonomous tasks, it is necessary to further clarify user behavior standards and usage boundaries.
On October 6, OpenAI publicly disclosed 722 mathematics research manuscripts generated by internal, yet‑unreleased cutting‑edge AI models, covering 372 research results across geometry, algebra, theoretical computer science, and other fields.
However, on October 8, OpenAI withdrew three of those papers. According to the project revision log, one paper contained a sign error that invalidated a key proof step, and the other two were affected because they relied on the related mathematical construction. OpenAI also revised another 14 papers, addressing proof fixes, conclusion corrections, clarification of assumptions, and citation updates.
At the same time, it is worth noting that UniPat AI recently released the scientific paper reproducibility benchmark PaperBenchX, which includes 93 real‑paper reproduction tasks from 12 research areas, covering the three stages of experiment modeling, execution, and result verification. The benchmark evaluates 10 cutting‑edge AI agent configurations. Among them, the GPT‑6 Astra paired with Codex ranked first, but its full reproducibility rate was only 13.98%—i.e., it successfully completed 13 out of the 93 tasks; its partial completion score was 59.08%. The benchmark treats task scores above 90% as full reproduction and assesses results by re‑running experiments and verifying evidence.
On October 7, Fields Medalist and UCLA mathematics professor Terence Tao reposted on his personal blog a statement issued by the Association for Human Mathematics (AHM), criticizing OpenAI for disregarding the norms of mathematical research and calling on mathematicians to stop cooperating with it.
The AHM statement notes that the mathematical community never asked OpenAI to conduct such research. Previously, the Advisory Group on Mathematics and Artificial Intelligence (AGMAI) had recommended that leading AI companies should not use internal models to test advanced mathematical problems. The AHM argues that OpenAI’s latest release ignores this recommendation and fails to follow academic standards that safeguard the credibility, ethics, and public interest of mathematical research.
AHM also questions OpenAI’s method of mass‑releasing mathematical results, arguing that publishing more than 700 documents at once is not a showcase of academic research but a display of corporate power. The organization states that it does not agree with OpenAI’s claim that this release advances the field of mathematics, and calls on mathematicians and the public to view this release model with caution. The statement concludes by urging mathematicians to cease collaborations with OpenAI and return to a science research approach centered on human understanding.
On October 5, media reports said that Meta and Microsoft are reducing employees’ use of Anthropic’s Claude model and related tools, and are pushing internal AI applications toward in‑house models and tools.
Earlier this year, Microsoft said it expected its internal spending on Anthropic technology to reach at least $1 billion annually. Since then, Microsoft has cut that budget by more than a third and asked employees to reduce their use of Claude in favor of its own AI tools.
At Meta, the number of employees using Claude Code internally has fallen from about 60,000 earlier this year to roughly 30,000 recently. Part of the decline is tied to previous layoffs, but a major factor is Meta’s push of its own AI coding tools and the Muse Spark series of models. Currently, more than 6,000 internal users are using Meta’s home‑grown coding tool Muse Code, and another internal‑only tool, MetaCode, has over 30,000 users.
The question of whether Microsoft’s and Meta’s reduction in internal Claude usage will impact Anthropic’s revenue has drawn market attention. After all, Anthropic is preparing for an IPO, and reports indicate that the two companies together accounted for about a quarter of its revenue in 2025.
However, Anthropic’s enterprise customer base is expanding. According to The Information, the company disclosed to investors that in the 12 months ending June this year, about 6,000 customers spent $100,000 or more on its AI software—four times the total number of customers for the entire year of 2025. Among them, over 1,000 customers spent more than $1 million, and more than 100 customers spent over $10 million.
At the same time, two tech giants remain important customers for Anthropic. Microsoft’s spending on Claude services delivered to external customers through Copilot continues to grow. Meta still pays Anthropic roughly $105 million per month. The “two customers contribute about a quarter of revenue” figure reflects the situation in 2025. Based on the roughly $4.6 billion revenue that year, the two companies together contributed about $1.1 billion. This historical amount represents less than 2% of Anthropic’s current annualized revenue scale.
On October 4, DeepSeek Harness team member Cui Tianyi responded to the introduction of a Claude Code Mods compatibility layer in the latest project version. He said that DeepSeek Harness’s core philosophy is “everything is a plugin,” a design principle that was established before the project was launched and before writing the first line of code, and extensibility was also an original design goal of the project.
Regarding the Claude Code Mods compatibility layer added in version v0.2.1-alpha.1, Cui Tianyi stated that this feature is still experimental and is mainly used to verify whether the extended capabilities that Claude Code Mods offers to plugin developers are essentially covered by the existing plugin architecture of DeepSeek Harness. Currently, this compatibility layer cannot support seamless operation of all Claude Code Mods. Cui believes that if development work continues in the future, a more complete compatibility could be technically achievable.
On October 9, the China National Intellectual Property Administration announced an appearance design patent for a embodied intelligent robot, with BYD Co., Ltd. as the patent holder. Public documents show a humanoid robot design featuring a black‑and‑white color scheme, a rounded curved head, and multiple articulated joints in the torso, arms, and legs; it is described as an intelligent robot capable of walking, balancing, and interacting, with applications in service, education, and industrial scenarios.
This patent disclosure provides new public information for observing BYD’s robot strategy. Previously, BYD Executive Vice President Li Ke discussed plans to deploy robots in stores to introduce and demonstrate vehicles to customers.
On October 9, Business Insider reported that Google employees recently began testing the next‑generation Gemini 4 model codenamed Carbon. An employee involved in the testing said Carbon’s programming performance is comparable to Anthropic’s Claude Opus 5.5, but further testing is needed.
The report said Google is simultaneously advancing several internal versions of the Gemini 4 series, including Argon, Barium, and Carbon. The version internally codenamed Barium‑B was selected as the publicly released Gemini 4 Argon, while Carbon is a newer internal test version.
Earlier, Google announced on October 1 the launch of Gemini 4 Argon, focusing on tasks such as complex reasoning, software engineering, and cybersecurity defense, and initially opened it to select cybersecurity partners through the Fairwind program, with plans to gradually expand its availability.
Currently, Carbon has not been officially released to the public, and Google has not disclosed its model parameters, benchmark results, or specific launch date. At present, descriptions of its programming capabilities mainly come from internal employee feedback, with no publicly available evaluation data. Google declined to comment on the internal tests or release plans.
On October 7, Anthropic released Claude Haiku 5.5, positioning it as a small model for high‑frequency, cost‑sensitive tasks. According to the official description, the new model is primarily intended for fast, repetitive work such as summary generation, context compression, database queries, and classification, and it can also act as a sub‑agent in programming workflows, collaborating with Opus 5.5 and Sonnet 5.5 to complete tasks. Real‑time customer service and browser automation—scenarios that are sensitive to response speed—are also key application areas.
Anthropic states that, compared with Haiku 4.5, Haiku 5.5 reduces average operating costs by about 75%. This figure is an average cost comparison provided by the vendor and does not mean that every task’s final bill will drop by the same proportion; actual costs still depend on input/output size, cache usage, and the way the task is executed.
This release also includes other price adjustments: the cache read cost for Sonnet 5.5 is halved, which the company says can reduce operating costs by about 20% for most agent tasks; Claude Max and Team subscription users will receive a new monthly API quota. Overall, Anthropic is lowering the barrier to entry across three fronts—smaller models, caching, and subscription benefits—to encourage developers to offload more high‑frequency tasks to different models.
On October 8, OpenAI announced the launch of the GPT‑6.1 Sol Ultrafast mode, now available in the API, Codex, and ChatGPT Work. According to the official description, the Ultrafast mode can make token generation up to eight times faster than the Standard mode while retaining capabilities close to those of GPT‑6 Astra. It is primarily aimed at scenarios that require high response speed, such as code debugging, AI‑agent‑driven application operations, and real‑time interaction.
In terms of pricing, the Ultrafast mode API costs $12 per million input tokens and $60 per million output tokens, which is six times the rates of the Standard mode. Developers can call the gpt-6.1-sol via the Responses API and set service_tier to ultrafast to enable this mode. OpenAI recommends using WebSocket in Agent applications that require continuous tool calls to reduce latency caused by network communication.
Currently, all API users can use Ultrafast mode under the applicable rate limits. Access in Codex and ChatGPT Work is limited to Pro 500 users, eligible pay‑as‑you‑go Enterprise customers, and Edu plan users based on credits, with enterprise admins needing to enable it proactively. This mode is now supported in all globally available regions and offers data residency options in the United States and the European Union.
On October 9, Microsoft officially launched the AI model Microsoft-Decision-1 for rapid decision‑making scenarios. It is currently available in open preview on Microsoft Foundry, and will later be offered through OpenRouter. The model is primarily aimed at tasks such as classification, routing, priority ranking, result verification, and workflow control within AI agents and software pipelines.
Technically, Microsoft-Decision-1 is fine‑tuned from Qwen3.5‑9B and performs decision scoring with a single forward pass. Unlike large language models that generate natural language text, this model outputs calibrated probability scores for predefined options, supporting binary judgments, multiple‑choice, rating, as well as scoring AI responses and agent behaviors.
In terms of performance, Microsoft compared several large language models and decision models across 36 benchmark tests covering roughly 150,000 questions. According to the official results, Microsoft-Decision-1 achieved the highest accuracy, reduced latency by about 78% compared to Quyet-1.0‑Large, and runs roughly 35 times faster than GPT‑6 Sol. Microsoft also demonstrated the model’s use in fault response and scientific experiment planning. Currently, its API is priced at $0.042 per million input tokens, with output tokens free of charge.
On October 8, Google Cloud unveiled the general‑purpose office AI agent Gemini agent at the Gemini at Work 2026 event, supporting knowledge work, content generation, code writing and execution, among other tasks. Users can assign work objectives through a unified conversational interface, and the agent autonomously plans and invokes tools to complete the tasks.
Unlike tools that rely on local devices, the Gemini agent runs tasks in the cloud, continuing for hours or even days after a user shuts down their computer. It supports scheduled jobs, event triggers, and cross‑device access, and can retain task context, memory, and personalized settings.
Gemini Agent also supports dynamically creating multiple sub‑agents to collaboratively handle complex tasks. Google also introduces the Coworker Agent, which enterprises can assign an independent email address, persistent storage space, and access permissions, allowing it to participate in long‑term work as a team member.
In terms of models, Gemini Agent is not limited to Google‑developed models; it can currently invoke Gemini series models as well as Anthropic’s Claude models depending on the task. The product can connect to services such as Gmail, Drive, Docs, Microsoft 365, and Slack. Currently, Gemini Agent is available in a private preview for enterprise customers and has not been fully released.
On October 6, Google DeepMind released the image generation and editing model Nano Banana 2.1, with model ID gemini-nano-banana-2.1. It is currently available through the Gemini API and related Google Cloud platforms. This model is an upgraded version of Nano Banana 2 (Gemini 3.1 Flash Image), continuing the Flash-level generation speed and cost positioning, with major improvements in image quality, prompt adherence, character consistency, and text rendering capabilities.
In terms of image generation, Nano Banana 2.1 supports output at 1K, 2K, and 4K resolutions, improves stitching artifacts in widescreen and panoramic ratios, and enhances the accuracy of infographic layout and text rendering. For multi‑image editing, the model can fuse up to 14 reference images, maintain consistency for up to 4 characters, and preserve the appearance features of up to 10 objects.
Additionally, the new model supports retrieving reference information by combining Google web search and image search, and offers three configurable reasoning levels: minimal, medium, and high. Inputs support text, images, video, and PDF, and outputs include images and text, with a context window of 131,072 tokens.
On October 8, Tencent’s AI office product WorkBuddy officially launched the “standalone file browser” feature, enabling users to open local files on their computer directly and invoke AI for summarization, analysis, editing, and polishing, without needing to upload the file to the AI chat window first.
According to the official description, users can right‑click a file in Windows File Explorer or macOS Finder and choose to open it with WorkBuddy. The new interface uses a layout that shows the file content on the left and provides an AI chat on the right, supporting file formats such as Word, Excel, PPT, PDF, Markdown, HTML, and more. Users can also set WorkBuddy as the default opening method, allowing them to double‑click a file to go directly to the AI‑assisted editing interface.
In addition, the new feature supports switching between multiple file tabs, with each file having its own independent AI conversation. Users can view and edit content in the same window and save the changes back to the local file.
Regarding privacy, Tencent states that the standalone file browser only reads files that users explicitly open or add, and does not traverse or scan the computer’s disk. Viewing or editing local files does not consume credits; invoking AI processing tasks consumes credits based on actual usage. Currently, users can use this feature as soon as they update WorkBuddy, without needing a separate download or installation.
On October 9, Doubao Work announced an update: the task mode now includes a “Creative Canvas” feature, and it also integrates the lightweight Doubao 2.1 Lite model and the image‑generation Seedream 5.0 Flash model.
According to the announcement, the newly launched Canvas feature is aimed at complex content creation tasks. Users can gather reference materials, design plans, images, videos, HTML, and other content onto a single infinite canvas for viewing, comparison, and organization. After generating content, they can further adjust text, colors, layout, and local details, and edit images using layer editing, cut‑out, erasing, and other tools.
In terms of usage, users can select “Create Canvas” in the work task mode, or generate content through skills such as “Creative Design” and “Creative Video.” The resulting outputs will be automatically added to the canvas. This feature will be gradually opened starting October 9.
Regarding models, Doubao 2.1 Lite is primarily aimed at everyday Q&A, document writing, spreadsheet handling, and PPT creation in office scenarios. The vendor claims it has been optimized for response speed and quota consumption, and supports image, video, and audio comprehension. The newly integrated Seedream 5.0 Flash is used for image generation and creative design tasks.
On October 6, French AI company Mistral released the public preview of Mistral Large 4, which developers can try via the Mistral Studio API. The company positions it as a trillion‑parameter‑scale, native multimodal mixture‑of‑experts model, focusing on programming, agent workflows, multimodal understanding, as well as enterprise tasks such as finance, law, and cybersecurity. Its documentation lists a context window of 1 million tokens.
Mistral says the model was trained from scratch in its own European data center using 3,800 NVIDIA Grace Blackwell GPUs, and the public preview is also served by that infrastructure. The company plans to open the model weights by the end of October and will simultaneously disclose more details about the architecture, benchmarks, and fine‑tuning; until then, testing will continue with vetted partners and relevant institutions.
On October 6, Google released EmbeddingGemma 2, an open multimodal embedding model with 740 million parameters. Unlike conversational models that generate answers directly, an embedding model converts content into vectors that can be compared and retrieved. EmbeddingGemma 2 supports text, code, images, video, and audio, mapping these inputs into a unified 768‑dimensional vector space, providing foundational capabilities for cross‑modal search, content retrieval, and retrieval‑augmented generation applications.
One of the highlights of this update is on‑device deployment. Google says that after quantization on the Pixel 11 Pro, the active memory required for the text‑only weights is as low as about 191 MB, while the full multimodal model requires roughly 567 MB.
Google also provides development guides and adds media search and video clip localization demos in the AI Edge Gallery. Developers can build image, audio, or video retrieval capabilities around natural language queries without having to build separate, disconnected retrieval systems for each type of content. For applications that require on‑device data processing, the unified embedding model also offers a way to reduce cloud transmission.
On October 9, Cloudflare and Deno jointly announced that Ryan Dahl, the creator of Node.js and founder of Deno, will lead the entire Deno team to join Cloudflare. The two parties plan to integrate the open‑source projects workerd and celld, enabling developers to run Cloudflare Workers and Durable Objects on their own infrastructure and support distributed scaling.
This integration will adjust Deno’s existing product roadmap. According to the official announcement, Deno Runtime will continue to receive monthly updates for the next year, providing bug fixes and security patches; after that, the team will cease development, but the project will remain open source. The cloud‑hosted platform Deno Deploy will be shut down in six months, and paying customers will receive support for migration to Cloudflare Workers. The JavaScript and TypeScript package registry JSR will continue to operate, with its underlying infrastructure migrated to Cloudflare.
Technically, Cloudflare has already open‑sourced the Workers runtime workerd, but its Durable Objects implementation still lacks the full distributed scaling capabilities required for independent deployment. The celld project developed by the Deno team is built in Rust, uses object storage as its primary external dependency, and provides distributed computing, persistent state, and real‑time communication capabilities.
Going forward, Ryan Dahl and Bert Belder will help drive the integration of workerd and celld, aiming to enable developers to deploy distributed applications compatible with the Workers programming model without relying on the Cloudflare hosting platform.
On October 7 local time, Microsoft announced the pre‑order and launch schedule for the Surface Laptop Ultra and opened pre‑orders for the Surface RTX Spark Dev Box. The Surface Laptop Ultra starts at $2,599 and is slated to launch on October 16; the Dev Box is priced at $5,999, with the first wave available for pre‑order through the U.S. Microsoft website and scheduled for delivery in November.
The two devices are both built around NVIDIA RTX Spark platform. Microsoft disclosed that the Surface Laptop Ultra’s top configuration includes a 20‑core Grace CPU, a 6144‑core Blackwell RTX GPU, and 128 GB of unified memory, targeting AI development, content creation, and professional computing. The company states it can run models with more than 120 billion parameters locally. The Surface RTX Spark Dev Box comes with a pre‑configured Windows development environment, ready for development out of the box. The device is pre‑installed with mainstream development tools, runtimes, and environments such as Visual Studio Code, Git, GitHub CLI, GitHub Copilot, WSL, Python, Node.js, and others.
Microsoft also introduced Windows’ “mixed intelligence” direction, combining local models, cloud capabilities, and agent runtime environments. The accompanying Microsoft Execution Containers are used to limit the scope of agent execution, while the Dev Box comes pre‑installed with various development tools. For developers, this release further connects local inference, model experimentation, and the everyday Windows work environment.
On October 6, GitHub published a technical article disclosing a Git infrastructure overhaul aimed at supporting the development of intelligent agents. The company noted that developers and a large number of agents operating code repositories simultaneously are changing the platform’s load characteristics: by September 2026, the number of commits on GitHub reached 7.38 billion, five times the amount a year earlier. Frequent commits, concurrent writes, and the massive reads triggered by CI are causing traditional scaling methods to hit new bottlenecks.
GitHub explains that the current architecture stores full repository copies on the local disks of multiple file servers. Adding replicas can improve read capacity, but it also increases the cost of write coordination. The new approach separates persistent storage from compute, with Azure Blob Storage holding the authoritative data, compute nodes handling caching and serving requests, and moving tasks such as compression and garbage collection out of the online request path.
According to internal benchmark results released by GitHub, the new architecture can increase write throughput by up to 35×. GitHub says the refactoring is underway, will be rolled out gradually while the platform remains operational, and will retain existing branch protection, code review, and auditing mechanisms.
On October 9, research firm Omdia released a report stating that global shipments of desktops, laptops and workstations in Q3 2026 totaled 58.1 million units, a year‑over‑year decline of 21.2%. Among them, desktop and desktop workstation shipments were about 11.7 million units, down 23.5% YoY; laptop and mobile workstation shipments were about 46.4 million units, down 20.6% YoY.
Omdia linked the market slowdown to factors such as early procurement, end‑user price hikes and channel inventory clearance. The firm noted that memory and storage prices continued to rise, and supply issues for processors, graphics cards and other chips also increased procurement difficulty for OEMs. Its estimates show that the share of DRAM and SSDs in PC costs has risen from the usual roughly 15% to nearly 40%, and multiple rounds of price increases by manufacturers have not fully offset upstream cost pressures.
In terms of vendor rankings, Lenovo remained first with about 14.9 million units shipped, followed by HP, Dell, Apple and ASUS. Omdia also forecasts that global PC shipments in Q4 2026 could fall 24% YoY, with a further 7% decline for the full year 2027.
On October 5, Microsoft’s Windows 11 operating system marked its fifth anniversary since official launch. The system was released on October 5, 2021, and its global user base has now exceeded 1 billion.
Microsoft disclosed in the FY2026 Q2 earnings call that Windows 11’s user count has reached 1 billion, a year‑over‑year increase of over 45%. It took Windows 11 1,576 days to reach 1 billion users from launch, 130 days fewer than the 1,706 days Windows 10 needed to achieve the same scale.
In October 2025, Windows 10 ends regular support, and Microsoft continues to push users to upgrade to Windows 11. This October, Microsoft also announced advancing the Windows hybrid intelligence architecture, supporting AI agents to invoke local or cloud models based on task requirements, and introducing a security isolation mechanism for agents.
On October 4, Linux founder Linus Torvalds announced the release of the Linux 7.3‑rc6 test version, noting in his mailing list that with AI and large language models participating in code review and bug fixing, Linux kernel development is entering a new “AI normal.”
Torvalds said that this week’s code changes and commit count have returned to normal levels, with about half of the modifications coming from driver fixes covering multiple subsystems such as GPU, networking, USB, TTY, and audio. However, he specifically noted that the current “normal” commit volume includes a large number of small fixes to error‑handling paths, which is related to AI tools being involved in kernel development.
Previously, several release candidates of Linux 7.3 already featured issues discovered or patches generated with AI assistance, and related fixes continued through the rc3 to rc5 stages. In addition to driver updates, Linux 7.3‑rc6 also includes fixes for KVM virtualization, core networking, SMB file‑system client, BPF, and adds numerous kernel self‑tests.
Regarding timing, Torvalds said that the upcoming Linux kernel maintainer summit may affect the next version’s code statistics, but it is not expected to change the overall release schedule. Based on the current development cycle, the Linux 7.3 stable release is expected in mid‑to‑late October.
On October 4, Huawei executive director and head of the Terminal BG, Yu Chengdong, posted a video stating that Huawei Semiconductor has successfully designed and mass‑produced 381 Tao (τ) chips in smartphones, AI, general computing, networking and smart automotive sectors.
In May this year, Huawei director and president of the Semiconductor Business Unit, He Tingbo, introduced the “Tao (τ) Law” for the first time at the 2026 International Conference on Circuits and Systems, advocating “time scaling” instead of the traditional “geometric scaling,” improving chip and system performance and energy efficiency by reducing data transfer and computation wait times.
According to Yu Chengdong, as transistor sizes approach physical limits, relying solely on advanced process nodes to boost chip performance is becoming challenging. Huawei proposes a core strategy based on technologies such as Logic Folding, performing collaborative optimization across four levels—device, circuit, chip, and system—including shortening signal transmission paths, optimizing circuit layout, and reducing computation and communication latency through hardware‑software co‑design.
Based on Huawei’s disclosed roadmap, it is expected that by 2031 the transistor density of high‑end chips, following Moore’s Law, could reach a level comparable to a 1.4‑nm process.
CLAUDE.md or AGENTS.md? Anthropic’s answer: you won’t need either in the future
The annual AI summit is scheduled for 10.29, inviting you to join online.
The 2026 Tencent Global Digital Ecosystem Conference is about to commence. Main summit + International Expansion Summit + 40+ themed sessions, covering AI trends, industry implementation, and functional efficiency across three dimensions.
General office scenarios: meeting minutes, invoice verification, weekly report generation, customer service Q&A—AI saves you time. Enterprise application scenarios: contract review, resume screening, report verification, seal scheduling—AI reduces your costs.
Over 100 new AI product releases, 500+ Agent case studies covering all scenarios. Prosumer seven-function special session, with hands‑on demos for every role. Live online throughout, click 【Read the original】 to reserve and watch.
Get MirrorSpeed