OpenAI unveils an AI solution for the Navier–Stokes equations
要闻
OpenAI announced that its internal multi-Agent system produced a solution to the Navier–Stokes Millennium Prize Problem and released a proof and Lean formal proof; about 10,000 Agents worked for roughly 88 hours, followed by another 17 hours for verification.
OpenAI has published its internal multi-Agent system’s solution to the Navier–Stokes existence and smoothness Millennium Prize Problem, together with a proof manuscript and Lean formal proof. According to the company, the proof constructs a three-dimensional incompressible fluid that is initially smooth and stationary. Under a smooth external force, its velocity tends to infinity within a finite time while its total energy remains finite throughout, thereby establishing cases C and D in the official problem statement. In the construction, vortices continuously contract inward, stretch along the axis, and accelerate.
The solution used a next-generation internal model that remains in training and, according to OpenAI, significantly exceeds GPT-6 Astra in capability. About 10,000 Agents collaborated for roughly 88 hours to obtain the solution, after which GPT-6 Astra ran for an additional 17 hours to complete the Lean formalization and verification. The process generated approximately 2,700,000 messages and consumed about 130,000,000,000 output tokens. OpenAI said the system had previously produced a disproof concerning the regularity problem for the Euler equations without external forcing. OpenAI also acknowledged the priority of work by Levent Alpöge and Tristan Buckmaster on the Euler problem with external forcing and said it does not intend to claim the Millennium Prize associated with this result.
OpenAI has released ChatGPT Images 2.5 across ChatGPT platforms and via API. According to the company, the model cuts generation latency by up to 50% versus Images 2.0 while improving subject fidelity and localized editing.
OpenAI has launched its next-generation image model, ChatGPT Images 2.5, for all ChatGPT, ChatGPT Work, and Codex users across desktop, mobile, and web, while two related API models are also available. According to the company, the model reduces generation latency by up to 50% compared with Images 2.0, produces more natural lighting and textures, and better preserves people from reference photos. Multi-turn editing can modify only specified areas while leaving other details unchanged. ChatGPT has also added templates for Sketch drawings, posters, and other formats, along with image comment editing. Users can enter @Sketch in the chat box to generate a sketch image directly and can include the prompt when sharing an image. The API offers GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst. Flare is positioned as the default choice for most applications, while Sunburst targets high-end creative use cases with higher precision but longer generation times.
DeepSeek begins testing an interim version of V4.1 Flash
要闻
DeepSeek’s V4.1 Flash interim release has entered testing for API users without requiring changes to base_url. It is priced the same as deepseek-v4-flash and capped at 20 concurrent requests per account.
DeepSeek has opened testing of its V4.1 Flash interim release to API users. Users can keep base_url unchanged and set the model name to deepseek-v4.1-flash-expires-on-0910, with 0910 serving as the version number. Current pricing is the same as deepseek-v4-flash, and the rate limit is 20 concurrent requests per account. According to the company, the release uses a new model architecture, supports multimodality natively, and delivers stronger capabilities, higher speed, and lower cost. DeepSeek is also providing an anonymous questionnaire that collects account information, the Agent tools or frameworks used, and testing feedback, while asking whether the model can fully replace the online DeepSeek V4 Pro.
DeepSeek to adjust pricing for the flash series starting at 12:00 on September 10
要闻
DeepSeek will revise pricing for its flash series at 12:00 Beijing time on September 10, 2026. The three off-peak unit prices will be RMB 0.02, RMB 1, and RMB 4, while peak-hour rates will be twice those amounts.
DeepSeek announced that new pricing for the flash series on its open platform will take effect at 12:00 Beijing time on September 10, 2026. According to the notice, the off-peak unit prices for input cache hits, input cache misses, and output will be RMB 0.02, RMB 1, and RMB 4, respectively, with the corresponding peak-hour prices set at twice the off-peak rates. Community users said the previous off-peak unit prices for the three categories were RMB 0.05, RMB 1.5, and RMB 4.5, producing calculated reductions of 60%, 33%, and 11%, respectively. The notice also reminded users to plan their usage appropriately.
小米 officially opens invitation-only testing for the MiMo Desktop client
要闻
Xiaomi has launched invite-only testing for Xiaomi MiMo Desktop, allowing approved applicants to try 2 Preview models for free under time and capacity limits. According to the company, the client can produce editable PPT files, webpages, 3D content, and Apps.
Xiaomi has officially opened invite-only testing for the Xiaomi MiMo Desktop client and the Preview versions of its new-generation models. Applicants must apply through the testing page and pass a review before downloading and using the client. According to the company, the application is designed for real work scenarios and can accept materials in multiple formats, interpret a goal, break it into tasks, invoke tools, and produce editable PPT files, webpages, 3D content, and Apps. It also supports interactive previews, localized edits, and version rollbacks. Approved applicants can try MiMo-X-Pro-Preview and MiMo-X-Flash-Preview for free during the testing period, subject to time and capacity limits, with priority given to existing users of the Xiaomi MiMo Open Platform. Xiaomi says the MiMo-X series remains in preview testing and that its capabilities may change with version updates. Computer-control features are limited to the overseas version, and the service is not currently available in the European Union, the United Kingdom, or South Korea.
Google DeepMind has launched AlphaGenome Atlas for the global scientific community, using 1 PB of data to cover all 9 billion possible single-letter DNA changes in the human genome while providing impact scores and access through the web and API.
Google DeepMind has opened AlphaGenome Atlas to the global scientific community. It is available through its website, the AlphaGenome API, and a skill in Google Antigravity, with a Google Cloud version coming soon. The database uses the AlphaGenome model to precompute the regulatory effects of all 9 billion possible single-letter DNA changes in the human genome, producing a 1 PB dataset. According to the company, this is more than 30 times the size of the AlphaFold database. It also provides an AlphaGenome Variant Impact score that combines features from AlphaGenome, AlphaMissense, and other sources to rank variants from lower to higher impact and show mechanisms such as disruption to gene switches or RNA splicing instructions. Sundar Pichai said the resource can be used directly in a standard browser without programming and is free for academic researchers.
The Ant Ling team has open-sourced the multimodal Ling-3.0-flash-VL model. Its BF16 and FP8 variants are now available to download, with FP4 and INT4 quantized versions to follow.
The Ant Ling team has now open-sourced Ling-3.0-flash-VL and made its BF16 and FP8 variants available for download. The FP4 and INT4 quantized versions are not yet available and, according to the team, will be released later. The model has been published under the inclusionAI account on Hugging Face and ModelScope, where the two available precision variants can be obtained. The team says Ling-3.0-flash-VL goes beyond visual recognition by following visual cues to process images, videos, documents, and UIs, perform reasoning, search, and verification, use tools, and check results before delivering its output.
Nex-AGI has released the Nex-N2.5 Agent models in three variants: mini, Pro, and Max. The weights are planned for open-source release alongside hosted services, while Max uses a 1.6-trillion-parameter text-only MoE model.
Nex-AGI formally introduced the Nex-N2.5 Agent model family through its project announcement and said it will open-source the model weights while also providing hosted online services. mini and Pro build on Nex-N2’s multimodal foundation, with improvements focused on computer use, web browsing, and visually grounded Agent capabilities. Max is based on a 1.6-trillion-parameter, text-only Mixture-of-Experts (MoE) foundation model. According to the company, this is Nex-AGI’s first complete post-training effort at the trillion-parameter scale. The company says the models can continuously operate computers and browsers, autonomously execute and test programs, and use visual feedback to verify results and self-correct.
The evaluation suite covers coding, agentic workflows, computer use, and multimodal understanding. Nex-AGI’s evaluations use temperature = 0.7, top_p = 0.95, and top_k = 40, with the NexAU harness for coding tasks and the NexCUA harness for computer-use and browser-use benchmarks. For deployment, the company provides the Docker image nexagi/sglang:v0.5.18-nex-patch with its customized sglang fork preinstalled. Setting reasoning_effort to "medium" in an OpenAI-compatible Chat Completions request enables adaptive thinking, while function calling requires adding --tool-call-parser qwen3_coder when launching the server.
OpenUI released OUI-1, an open-weight model that generates user interfaces in openui-lang. According to the company, it can run in FP8 on an RTX 5090, while its Generative UI benchmark score rose from 13.0% to 57.1%.
OpenUI has released the open-weight OUI-1 model, with its weights available on Hugging Face under the Gemma Terms of Use, for generating user interfaces in openui-lang. According to the company, the model is fine-tuned from DiffusionGemma and can run locally in FP8 on an RTX 5090. Unlike an autoregressive model that generates one token at a time, DiffusionGemma processes a 256-token block at once. OpenUI Lang uses up to 67% fewer tokens than JSON and can stream an interface before the model finishes generating it.
Training consisted of two stages: supervised fine tuning and self-distillation. The initial stage used about 700 OpenUI Lang samples generated by larger models across 7 component libraries and performed a LoRA fine-tune on one A100. After narrowing the scope to the single component library used by the benchmark, the score rose from 13.0% to 28.8%, but generation time on the same 20 light briefs increased from 1.6 seconds to 4.3 seconds per output. The model then generated several hundred programs, which were accepted by a parser or repaired according to reported errors before a judge checked whether they matched their briefs. Accepted samples were used for 500 training steps, with each round taking 1–2 hours on one A100. According to OpenUI’s published results, self-distillation reduced generation time to 1.9 seconds while producing 28% more tokens than the base DiffusionGemma model, and the score reached 57.1%. Schema errors fell from 292 to 76, while wiring errors fell from 971 to 484. The process was subsequently expanded to 27 component libraries.
Inception Labs released Mercury 2.5, which it says improves intelligence by 40% over Mercury 2, runs at 1,107 tokens per second, and supports a 260K-token context window.
Inception Labs has officially released Mercury 2.5 and made it available through the Inception API, OpenRouter, and Baseten, with 100 million free tokens included for new accounts. According to the company, the diffusion LLM is its strongest production model to date and the strongest diffusion LLM on the market. Official data shows that its intelligence level is 40% higher than Mercury 2 and comparable to GPT-5.6 Luna (Low) and Gemini 3.5 Flash-Lite. It runs at 1,107 tokens per second on widely available NVIDIA GPUs and supports a 260K-token context window. Its features include adjustable reasoning, parallel tool calling, and schema-aligned JSON. Previews of Mercury Voice and Mercury Router are also available. The company said training of its next-generation model has begun, with release planned within several months.
Meta officially launches Muse, a personal AI Agent app
产品应用
Meta has launched Muse, a personal AI Agent app that can handle multiple tasks in parallel and continue working in the background after the app is closed. According to the company, it has its own computer, file system, terminal, and full web browser, allowing it to fill out forms, make bookings, and complete purchases.
Meta has officially launched Muse, a personal AI Agent app designed to pursue goals and execute tasks proactively while continuing background work after the app is closed. According to the company, Muse has its own computer, file system, terminal, and full web browser, enabling it to write code, build tools required for a task, search and navigate websites, fill out forms, and complete bookings and purchases. It can also generate documents, PDFs, and web pages. Users can submit multiple tasks at once or interrupt an ongoing process. Muse continues working based on schedules or relevant events, then determines whether there is new information or a need for user input before sending a notification.
The product centers on one continuous main chat, without requiring users to wait for a response to the previous task before sending another instruction. Its memory persists across conversations, while complex topics can be separated into side chats. Muse can proactively send messages without a new prompt, and users can turn this function off, dial it down, or dial it up. Users can also create an avatar, give it a name, and set its style. To expose background activity, the interface shows a work summary beneath the avatar; tapping it opens the complete activity log and approved permissions. Settings and Memory files can be viewed or edited directly, while the Goals tab lists the items Muse is tracking and its execution plans.
DaVinci Resolve 21.1 released with Claude and other AI assistants integrated
产品应用
Blackmagic Design’s DaVinci Resolve 21.1 launched on September 8, 2026, with support for project tasks using AI assistants such as Claude. Advanced Python scripting has moved to the Studio edition.
Blackmagic Design released DaVinci Resolve 21.1 on September 8, 2026, and made it available as a free download from the company’s website. The company said the update integrates AI assistants and expands camera format support. According to media reports, the new version can work with Claude, Claude Code, ChatGPT Codex, and other assistants to perform project analysis, media management, batch rendering, and highlight editing from long-form videos through prompts. The official release notes also state that advanced Python scripting has moved from the free edition to the Studio edition, while Windows and Linux support is limited to a single processing GPU.
Runway launches an Adobe plugin for generating videos directly in Pr and AE
产品应用
Runway has released an Adobe plugin that lets users generate images and video, restyle footage, and insert results within Premiere Pro and After Effects timelines. The plugin is free to download, while generation is available on all paid plans and consumes credits.
Runway has released Runway Plugins for Adobe Premiere Pro and After Effects. Users can generate images and video, restyle footage, and place results directly into a comp or sequence without leaving the timeline. According to the company, the plugin panel removes the process of exporting a frame, uploading it to a browser, downloading the result, and importing it again. Generation progress appears in the feed, and in Premiere Pro, results are placed at the playhead and can be edited immediately.
Edit Studio can process a clip selected in a comp or sequence. Runway says that after users restyle the anchor frames, Aleph 2 re-renders the entire clip at the same duration as the source, allowing the before-and-after versions to be compared in the panel before the result is returned to the edit. Users can sign in with an existing account and switch teams through a workspace switcher, while generations use their existing plan credits. HDR conversion, background removal, and upscaling can be applied with one click to any selected clip or feed asset. The plugins are free to download for macOS and Windows, and in-plugin generation is available on all paid plans.
OpenAI's Navier-Stokes proof sparks authorship and data controversy
行业动态
Mathematician Tristan Buckmaster said OpenAI pursued a related proof after learning of his research progress and sought to exclude his collaborator because he worked at Anthropic. OpenAI denied viewing the pair’s unpublished work while solving the problem.
Mathematician Tristan Buckmaster disclosed in a public statement that he and OpenAI had disputed the publication arrangements, paper authorship, and use of Codex data surrounding a Navier-Stokes-related proof. According to Buckmaster, OpenAI began related research after learning about progress made by him and collaborator Levent Alpöge, then proposed that Buckmaster alone write a paper presenting OpenAI’s proof because Alpöge worked at Anthropic and therefore would not be listed as an author. Buckmaster said that after rejecting the proposal and telling the other party he would make the events public, he received remarks concerning his career. He also asked whether drafts the pair uploaded to Codex had been used for model training, while stating that he did not know whether the data had actually been used.
OpenAI said in an announcement that the research was prompted by related rumors, that the company had originally hoped to coordinate publication, and that it recognized Buckmaster and Alpöge’s prior work on the forced Euler equations. According to the company, its researchers and Agent had not seen the pair’s work before their results became public and had not accessed specific user data to solve the problem. It said, however, that it could not rule out the possibility that de-identified data generated when the pair used its products had helped improve the model. OpenAI researcher Sébastien Bubeck said the authorship discussion concerned a new paper presenting OpenAI’s proof, not a request to remove Alpöge’s name from his own research; he also apologized for wording about “destroying a career.”
Mistral raises €3 billion in Series D funding at a post-money valuation of over €21 billion
行业动态
Samsung Electronics led Mistral’s €3 billion Series D, giving the company a post-money valuation of more than €21 billion. The company says the transaction set a record for equity fundraising by a European technology company.
Mistral announced the completion of a €3 billion Series D at a post-money valuation of more than €21 billion, with Samsung Electronics leading the round. Scaleup Europe Fund, managed by EQT, and existing investor PSG Equity served as co-leads. Advent, funds and accounts managed by BlackRock, and the Grand Duchy of Luxembourg joined as new investors, while existing investors including a16z, ASML, NVIDIA, and Salesforce Ventures also participated. The round was completed three years after the company’s launch. According to Mistral, it was the largest equity financing ever completed by a European technology company; its previous Series C was led by ASML.
According to the company, the funding will be used to expand frontier research and compute capacity for model training, build out infrastructure, and advance commercial growth and international expansion. Mistral currently operates in 20 countries and supports mission-critical AI transformation for 125+ global enterprises, including Airbus, ASML, and HSBC. The company defines its offering as a full stack comprising open-weight models, infrastructure, compute capacity, and production-grade products. It says this structure prevents customers from being constrained by a single vendor’s roadmap, pricing, or availability, while avoiding the need to expose data, workflows, and institutional knowledge outside their organizations. What it calls sovereign AI covers four areas of control: keeping data within organizational boundaries, making models controllable and customizable, keeping compute private and predictable, and ensuring production systems are controllable and auditable.
Cognition announces the completion of a Series E funding round exceeding $2 billion
行业动态
Cognition has completed a Series E round of more than $2 billion at a $48 billion valuation; since its previous round, run-rate revenue has risen from $492 million to nearly $900 million.
Cognition announced the completion of its Series E round, raising more than $2 billion at a $48 billion valuation. New investors Andreessen Horowitz and Accel led the round, with existing investors Founders Fund, General Catalyst, and Avenir participating alongside firms including Benchmark, Bessemer, Kleiner Perkins, Greylock, and Lightspeed. The company disclosed that its run-rate revenue had increased from $492 million to nearly $900 million since its previous funding round. According to Cognition, Devin is used in chip design at NVIDIA, aviation at GE Aerospace, financial services at Citi, automotive at Mercedes-Benz, and AI infrastructure at Modal.
Cognition was founded in 2024 with engineers positioned as architects who set goals and priorities while delegating more execution work to agents. The company says its current features include Devin Auto-Triage for initial incident investigations, Devin Security Swarm for finding and triaging vulnerabilities, and Devin Automations for starting work from events in systems such as Slack, GitHub, and Linear. Cognition says that operating as an independent agent lab allows it to select and combine models suited to specific work, including its own models, without tying customers to a single provider. Over the past year, it has opened offices in Washington, D.C., Tokyo, Singapore, London, São Paulo, and Madrid, in addition to hubs in San Francisco, New York, and Austin.
OpenAI contracts dedicated computing capacity from two Firmus AI factories in Malaysia
行业动态
Firmus and OpenAI announced a multi-year agreement on September 8, 2026, under which OpenAI will secure dedicated compute from two Malaysian AI Factories, taking Firmus’ total contracted capacity across all customers above 900 MW.
Firmus and OpenAI entered a multi-year strategic partnership on September 8, 2026, with OpenAI securing dedicated AI compute from two Firmus AI Factories in Malaysia and becoming an anchor customer. Including this agreement, Firmus has more than 900 MW of contracted capacity across all customers. Its portfolio comprises seven AI Factories in four countries: Australia, Singapore, Indonesia, and Malaysia. Two facilities in Australia and Singapore are operating, while five remain under development and are targeted to reach ready-for-service status within the next 24 months.
The AI Factories covered by the partnership will use the NVIDIA DSX AI Factory Platform to integrate NVIDIA Vera Rubin NVL72 rack-scale systems with the Firmus HyperCube platform. HyperCube combines liquid cooling, mechanical systems, and electrification in a repeatable infrastructure system manufactured and prefabricated in regional New South Wales. According to Firmus, the company will deploy NVIDIA Vera Rubin at scale across Asia-Pacific, with the related facilities covering power, cooling, accelerated computing systems, networking, software, and workload performance. Firmus also plans to establish an Australian AI Access Program supporting eligible researchers and organizations that apply AI in science, education, agriculture, energy and resource efficiency, and climate resilience.
