NVIDIA Blog published this on Oct 19, 2026 (source: blogs.nvidia.com).
The European Broadcasting Union, an alliance of more than 115 public service media organizations, is widening the sovereign standards it introduced with NVIDIA last year so broadcasters can deploy AI agents while keeping control of their data, infrastructure, models and security. The initiative will define requirements and reference architectures and now includes Microsoft Sovereign Cloud with Azure Local and Foundry Local, Nebius for training and running models on European infrastructure, NVIDIA Nemotron open models, and integrators such as Qvest. At the same Berlin event, NVIDIA says Adobe Frame.io's GPU-accelerated video search, used by German producer UFA, generates embeddings about 5x faster with up to 4x higher throughput.
Inception published this on Oct 19, 2026 (source: inceptionlabs.ai).
Inception introduced Mercury 3, which it calls the most capable diffusion-based LLM released to date and more than 40x larger than Mercury 1. Inception says it is on par with or better than GPT-6 Luna and Claude Haiku 5.5 across a range of benchmarks while running 4 to 8x faster, at 1,068 tokens per second on NVIDIA GPUs, with a 1M-token context window, up from 260K for Mercury 2.5. It is priced at $0.20 per million input tokens and $0.70 per million output tokens, half off at launch, and is available through an OpenAI-compatible API on Inception's platform, OpenRouter, Baseten and Vercel. Mercury 2 and 2.5 are closed to new users, and existing customers keep access until November 15.
U.S. Department of Homeland Security (Federal Register) published this on Oct 19, 2026 (source: federalregister.gov).
The Department of Homeland Security published notice of an order excluding IT products and services from Hangzhou DeepSeek Artificial Intelligence and its affiliates from federal civilian agency procurements. The order also removes all DeepSeek and DeepSeek-affiliate products, plus third-party products and services built using DeepSeek implementations, from civilian agency IT systems. It was issued under the Federal Acquisition Supply Chain Security Act after a Federal Acquisition Security Council review, and the notice was filed on Oct. 19 for publication in the Federal Register on Oct. 20. CyberScoop reports that the signed order from Homeland Security Secretary Markwayne Mullin is dated June 26 and cites risks including data transfers to China.
Not confirmed by the company — reported on Oct 19, 2026 (source: theverge.com).
AI security firm Norforge says a hacker broke into more than 100 websites with AI agents that mostly worked on their own, and Reuters reports the victims included John Deere, SharkNinja and Cummins. Norforge found the activity through exposed instances of Cairn, an open-source agent platform, and counted more than 20 exposed instances that are likely malicious. The most active attacker targeted companies in the US, Canada, Australia and Japan, set out preferred intrusion tactics in under 100 Chinese characters, and took about 23 hours per successful breach, per Reuters. John Deere said there is no sign sensitive information was accessed, Cummins is investigating a possible attack on a third-party vendor's website, and SharkNinja said a standalone product-support site was affected and the threat is contained, per The Verge.
U.S. Department of State published this on Oct 19, 2026 (source: state.gov).
The State Department designated five China-based entities under the Protecting American Intellectual Property Act (PAIPA) in connection with theft of trade secrets from U.S. persons: Shanghai Zhisuan Technology, which it says was established by Linwei Ding in 2023; the venture firm MiraclePlus (Qiji Chuangtan Beijing Investment Management); and three MiraclePlus entities, its Phase I and Phase II venture funds and Zhiyuan Enterprise. The State Department says Ding, convicted by a U.S. jury on January 29, 2026 of stealing Google trade secrets related to super intelligence (SI) technology, pitched Zhisuan at MiraclePlus as able to replicate Google's SI infrastructure technology, and that MiraclePlus took an ownership stake in Zhisuan in exchange for seed capital. OFAC added all five to its Non-SDN Menu-Based Sanctions List: their property is blocked, U.S. persons may not buy significant amounts of their equity or debt, and U.S. bank loans, foreign exchange and banking transactions involving them are restricted. OFAC General License 2 allows wind-down transactions with the four MiraclePlus entities until 12:01 a.m. EST on January 19, 2027.
OfficialLangGraph CLI 0.4.37 makes serverless the default deploy target
GitHub langchain-ai/langgraph published this on Oct 19, 2026 (source: github.com).
LangChain released LangGraph CLI 0.4.37. The release notes say deploy now defaults to serverless, and verbose output no longer prints the stdin passed to subprocesses. The rest of the release is dependency bumps in the CLI's JavaScript examples.
OfficialGoogle says IDC named it a Leader in the first IDC MarketScape for agentic AI platforms
Google published this on Oct 19, 2026 (source: cloud.google.com).
Google Cloud says it was named a Leader in IDC's inaugural MarketScape: Worldwide Agentic AI Platforms 2026 Vendor Assessment. In its announcement, Google argues that the hard part for companies is deploying agents with quality, integration, security and governance, not raw model intelligence. It points to Gemini agents working inside Google Workspace, Microsoft 365 and team chats with shared memory and permissions, builder tools from no-code to the Agent Development Kit and Antigravity, shared skills in a central Agent Registry, and more than 200 models in Model Garden, including Gemini 4 Argon and Anthropic models.
Reported by TechCrunchSmart TV owners are using AI coding agents to strip ads and tracking from their TVs, per TechCrunch
Not confirmed by the company — reported on Oct 19, 2026 (source: techcrunch.com).
Developers and vibe coders are sharing AI-made fixes that remove sponsored ads, disable tracking features, speed up app launchers and redesign smart TV home screens, TechCrunch reports. One project, dumbtv, says a TV should run only your apps and HDMI inputs, with no ads, watch tracking or unrequested updates; another, cleanyourtv, strips bloatware from Android TV and Google TV over Wi-Fi and swaps in an Apple TV-style launcher. Its creator says Claude Code found his parents' Xiaomi TV on Wi-Fi, removed 25 packages and wrote a new launcher, and that a few thousand people did the same within two days. TechCrunch notes this could hurt the smart TV ad business, which sells viewing data gathered by Automatic Content Recognition, per TechCrunch.
Reported by The VergeIllinois Gov. JB Pritzker says AI may help his kids live to 100 but compares its risks to nuclear weapons, per The Verge
Not confirmed by the company — reported on Oct 19, 2026 (source: theverge.com).
On The Verge's Decoder podcast, Illinois Gov. JB Pritzker said he expects his children to live to 100 or 110 because AI will solve many medical problems, but said AI is potentially existential and that managing rivals may end up looking like nuclear-era mutually assured destruction. He said Illinois, New York and California are the only states with AI safety laws so far and that Congress's inaction leaves states going it alone. He also cited Illinois's ban on AI therapy, argued that the backlash against data centers is about electricity bills rather than AI, and said privacy fights such as those over Flock cameras are no longer left or right, per The Verge.
UnconfirmedUnannounced GPT-6.1 Terra shows up in a Microsoft pull request that switches 18 eval configs to it, LuminaBench says
Not verified by an official source as of Oct 19, 2026 (source: x.com).
The X account LuminaBench says a Microsoft pull request changed 18 evaluation configurations from GPT-5.6 Terra to GPT-6.1 Terra, a model OpenAI has not announced. Its post followed a note from The Information's Stephanie Palazzolo that OpenAI may be preparing a new Terra model, alongside a new Anthropic model and Google's Gemini 4 Carbon. OpenAI and Microsoft have not commented.
Reported by The Daily CalifornianFields medalist Richard Borcherds says human math research could be redundant in a few years after OpenAI's Navier-Stokes result, per The Daily Californian
Not confirmed by the company — reported on Oct 19, 2026 (source: dailycal.org).
UC Berkeley mathematician Richard Borcherds, who won the 1998 Fields Medal, told The Daily Californian he had not expected AI to reach the level of OpenAI's announced Navier-Stokes result for about 10 years. "It's fairly obvious that if things go like this, in a few years, humans doing math research will be redundant," he said, adding that it is hard to know what to tell graduate students. The paper notes that OpenAI says an internal model running about 10,000 agents produced the result, and that OpenAI researcher Noam Brown said it cost millions of dollars, per The Daily Californian.
OfficialTogether AI launches Together Post-Training, a service for RL and SFT on open models where users write the training loop
Together AI published this on Oct 19, 2026 (source: together.ai).
Together AI introduced Together Post-Training, a service for building custom models from open-weight LLMs with reinforcement learning from environment rewards or supervised fine-tuning. Users define the training loop themselves, including custom losses, LoRA or full-weight training for select models, and their own data, environments and agent harnesses, while Together runs the trainer and rollout infrastructure on dedicated GPU allocations billed per GPU-minute. It launches with Muse Spark Open, GLM 5.3, Kimi K3 and Qwen 3.8-27B, and Together offers a one-line migration for scripts written for the Tinker API. For the first eight weeks, Together is charging a reduced $9 per B200 GPU-hour.
OfficialHugging Face Transformers 5.20.2 fixes a generate bug that changed input types for multimodal models
GitHub huggingface/transformers published this on Oct 19, 2026 (source: github.com).
Hugging Face released Transformers 5.20.2, a patch release. The release notes say it fixes an issue where some inputs could lose their original type inside prepare_inputs_for_generation in generate, which caused problems with multimodal models. It also fixes zero expert gradients in bfloat16 with expert parallelism when ep_size is smaller than fsdp_size, declares the input modalities of vision-language models that reported text only, and avoids a host-to-device copy in eager_mask.
Reported by Upstarts MediaAI companion app Tolan raises $30M from USV, Khosla Ventures and Homebrew, per Upstarts Media
Not confirmed by the company — reported on Oct 19, 2026 (source: upstartsmedia.com).
Tolan, an AI companion app that helps people organize daily life and family tasks, has raised $30 million from Union Square Ventures, Khosla Ventures and Homebrew, doubling its total funding, Upstarts Media reports. The startup declined to share its valuation but said the round is a significant step up from its previous $120 million valuation, per PitchBook. Tolan says about 250,000 people use it, that about 70% of users are women and 40% are parents, and that annual recurring revenue has more than doubled from last year to the double-digit millions. It plans to use the money to widen access, including a first-ever free tier, and CEO Quinten Farmer says it is not yet profitable, per Upstarts Media.
Reported by TechCrunchReplika founder Eugenia Kuyda launches Wabi 2.0, a multiplayer AI assistant that builds apps inside shared chat threads, per TechCrunch
Not confirmed by the company — reported on Oct 19, 2026 (source: techcrunch.com).
Eugenia Kuyda, who founded Replika, launched Wabi for all users on Monday as a personal AI assistant that works like a messenger app, with separate agentic threads for different tasks and goals, TechCrunch reports. Once connected to a user's email, Wabi 2.0 spins up threads and builds custom apps inside them on its own, such as a weight-loss tracker in a health thread or a nearby kids' events app in a family thread. Threads can be shared as group chats, access to personal data such as calendars is scoped to each chat, and everyone in a chat can see when someone's personal information is requested. Wabi, which raised a $20 million pre-seed round last year, is free for now, has not settled on a business model and is in talks to raise more, per TechCrunch.
OfficialViggle unveils JST-2, a world model that generates explorable 3D scenes in about a second, plus rigged 3D characters and motion
Viggle published this on Oct 19, 2026 (source: viggle.ai).
Viggle introduced JST-2, which it calls a physics-native world foundation model that represents scenes as persistent atoms in 3D space instead of learning in pixel space. Viggle says JST-2 generates a 3D scene, decoded into Gaussian splats for browser rendering, in about one second on NVIDIA B200 at an inference cost below $0.01, and accepts up to 64 input views so a world can be expanded one view at a time. The same model creates 3D characters with motion built in from a single image or text, generates full-body motion from text or video, and reconstructs 3D environments from sparse real-world images. The speed and benchmark comparisons are Viggle's own.
OfficialHugging Face libraries now run natively on AWS Trainium through TorchNeuron, which enters public beta Oct. 20
Hugging Face Blog published this on Oct 19, 2026 (source: huggingface.co).
Hugging Face says Transformers, Diffusers, Accelerate, PEFT, TRL and Kernels now support TorchNeuron, a PyTorch plugin from AWS that adds Trainium as a native device with eager execution, torch.compile and PyTorch's distributed APIs such as FSDP and tensor parallelism. TorchNeuron becomes a public beta on Oct. 20, installable with pip install torch-neuronx. Hugging Face says 89% of the top 100 most-used text-generation architectures work out of the box, compared with 46% under Optimum Neuron v0.4.5, and it now recommends that Neuron users use its libraries directly instead of Optimum Neuron.
OfficialdoubleAI says its Maxwell code execution model made LTX-2.5 generate 4K video nearly 2x faster
DoubleAI published this on Oct 19, 2026 (source: doubleai.com).
doubleAI says Maxwell, its code execution model that rewrites a running program's binaries for the exact machine and data without changing the source code, cut LTX-2.5's 4K video generation from 250 seconds to 134 seconds per clip on two NVIDIA B200 GPUs, a 1.87x speedup with the same output. It says Maxwell made more than a dozen changes to LTX's already optimized pipeline, from how work is split across tiles and GPUs down to edits inside individual kernels, and moved CPU work onto the GPU so the pipeline no longer depends on host CPU speed. The figures are doubleAI's own.
Reported by The Wall Street JournalAnthropic's Long-Term Benefit Trust can seat most of its board, but its governing documents are still private ahead of the IPO, per The Wall Street Journal
Not confirmed by the company — reported on Oct 19, 2026 (source: wsj.com).
Anthropic's Long-Term Benefit Trust, a five-member body whose trustees hold no equity, can appoint four of the company's seven board members and dismiss them without cause, but the documents that govern it have not been made public, The Wall Street Journal reports. Anthropic must give the trustees five days' notice before issuing more than routine debt, selling half or more of the company, or transferring control of any AI model above thresholds its board updates each quarter; some legal experts say that amounts to a veto, while an Anthropic spokeswoman said it does not. It is still unclear when and how shareholders could override the trust, which has two vacancies and includes former Fed chair Ben Bernanke and chairman Neil Buddy Shah. OpenAI, also a public-benefit corporation, is controlled by the OpenAI Foundation rather than a trust, per The Wall Street Journal.
OfficialOllama supports MiniMax M3.1 Flash from day one, US-hosted with zero data retention
@MiniMax_AI on X published this on Oct 19, 2026 (source: x.com).
MiniMax says Ollama supports its new M3.1 Flash model from day one. MiniMax describes the Ollama offering as US-hosted with zero data retention and ready for agents.
OfficialNVIDIA signs an industrial AI declaration of intent with Germany's government and plans investments in four German startups
NVIDIA Blog published this on Oct 19, 2026 (source: blogs.nvidia.com).
German Chancellor Friedrich Merz and two federal ministries signed a Joint Declaration of Intent with NVIDIA at the Invest in Germany Summit in Berlin, agreeing to work on a joint initiative to advance AI models for physical AI. Merz and NVIDIA CEO Jensen Huang then discussed how Germany can apply AI to manufacturing and automotive and spread it across small and medium-sized companies. Huang also announced planned investments in robotics maker Agile Robots, hospital software startup Avelios Medical, audio-for-video AI startup Mirelo and satellite builder Reflex Aerospace, which NVIDIA says are part of a nearly $10 billion commitment to companies, universities and research institutions across Europe.
OfficialMeta publishes a technical report on Muse Glimmer, its 30B open-weight model for agentic tasks on consumer hardware
@AIatMeta on X published this on Oct 19, 2026 (source: x.com).
Meta published a technical report on Muse Glimmer, its 30B open-weight model built for agentic tasks on consumer hardware, covering the model architecture, distillation recipe and quantization. Meta says the model runs at practical speeds on a Mac or a PC with a single consumer GPU at 4-bit precision and leads comparably sized open models on tool use and multi-step planning. The comparison is Meta's own.
OfficialMeta details Prometheus, the gigawatt-scale Ohio supercluster that trained Muse Wonder at more than 85% goodput
Engineering at Meta published this on Oct 19, 2026 (source: engineering.fb.com).
Meta says the first phase of Prometheus, its gigawatt-scale supercluster in New Albany, Ohio, is complete and was used to train its Muse foundation models, including Muse Wonder. Meta says Muse Wonder training reached more than 85% goodput on pre-training runs and more than 90% on reinforcement learning runs, with a mean time between interruptions of about 47,000 GPU hours, since raised above 54,000 with fleet fixes. Prometheus spans hundreds of thousands of GPUs across multiple buildings, is planned to reach roughly 1 million GPUs, and will add a third NVIDIA platform based on Vera Rubin NVL72 soon.
OfficialNVIDIA runs a record 104-qubit fault-tolerant quantum computing simulation on Europe's JUPITER supercomputer
NVIDIA Newsroom published this on Oct 19, 2026 (source: nvidianews.nvidia.com).
NVIDIA researchers ran a 104-qubit simulation of a fault-tolerant quantum processor on JUPITER, Europe's first exascale supercomputer at the Jülich Supercomputing Centre, using NVIDIA GH200 superchips. NVIDIA calls it the largest comprehensive simulation of a fault-tolerant quantum computing system to date. The simulation validates magic state cultivation, an approach NVIDIA says needs 90% fewer physical qubits than previous methods, and shows how noise affects designs before hardware is built.
OfficialCognizant and Google Cloud launch a Gemini Frontier Capability Group backed by more than 18,000 Cognizant engineers
Cognizant published this on Oct 19, 2026 (source: news.cognizant.com).
Cognizant and Google Cloud expanded their partnership with a new Gemini Frontier Capability Group, backed by more than 18,000 dedicated Cognizant associates, to help enterprises move AI agents from pilot to production. The group will start with legacy modernization, developer experience, business process automation and cyber defense, using an agent architecture built on the Model Context Protocol and secured with Google Cloud's Model Armor. Cognizant says more than 55,000 of its associates completed its Build with Gemini workshop.
OfficialCognition adds an Efficiency Score, Devin Coach and efficiency-based spend approvals for enterprise Devin users
Devin published this on Oct 19, 2026 (source: devin.ai).
Cognition introduced three enterprise features for controlling Devin spend. Efficiency Score rates each engineer's recent Devin usage from 0 to 100 based on the productive engineering hours Devin estimates per unit of spend, and Devin Coach flags problems in a prompt before it runs, such as a vague request or an expensive mode for a simple task. Efficiency-based approvals raise limits automatically for engineers with a Healthy or Satisfactory grade, up to a ceiling the admin sets, while other engineers' requests go to an admin for review.
OfficialMicrosoft says frontier models can't be trusted inside companies and expands Project Perception to supervise enterprise AI agents
Microsoft published this on Oct 19, 2026 (source: blogs.microsoft.com).
Microsoft argues on its official blog that frontier models, regardless of origin, cannot be trusted, and that AI agents deployed inside companies should be treated as agentic insiders with security controls outside the model. It is expanding Project Perception, its agentic security system, to discover agents from identity, data, cloud, endpoint, network and runtime signals, give each an identity and permissions, compute their risk dynamically, and have supervising Guardian Agents respond in real time as risk rises. CEO Satya Nadella said the new capabilities help continuously assess agent risk, adapt permissions and contain threats.
OfficialMiniMax Code offers unlimited M3.1 Flash through Oct. 25 and 50% off a first month of M Plan through Oct. 31
@MiniMaxAgent on X published this on Oct 19, 2026 (source: x.com).
MiniMax says every MiniMax Code subscriber can use M3.1 Flash with no usage caps through Oct. 25 at 23:59 (UTC+8). New subscribers and M Plan upgraders get 50% off their first month through Oct. 31 at 23:59 (UTC+8), with standard pricing from the second month.
OfficialGitHub makes its Copilot model policy REST API generally available for enterprises, teams, apps and organizations
GitHub published this on Oct 19, 2026 (source: github.blog).
GitHub's GET and PATCH endpoints for retrieving and updating Copilot model policies are now generally available at four levels: enterprises, enterprise teams, enterprise apps and organizations. Enterprise policies can delegate model control to organizations in organization mode, or to teams and apps in enterprise teams mode.
Reported by BloombergNscale and Anthropic are in talks on a second $45 billion compute deal, per Bloomberg
Not confirmed by the company — reported on Oct 19, 2026 (source: news.bloomberglaw.com).
Nscale and Anthropic are in talks on another compute deal worth about $45 billion, Bloomberg reports, citing people familiar with the matter who cautioned that talks are ongoing and details could change. Anthropic agreed in August to spend $45 billion renting computing power from Nscale's West Virginia data center campus. Reuters, citing a person familiar, says the new deal would cover a second building at the Monarch campus delivering 460 megawatts with Nvidia Vera Rubin chips, and that Anthropic has a right of first refusal on a third. Both companies are pursuing IPOs, per TechCrunch.
OfficialOpenAI releases MultiLangBench and says its latest models are far better at COBOL, Fortran and other low-resource languages
OpenAI News published this on Oct 19, 2026 (source: openai.com).
OpenAI introduced two evaluations for low-resource programming languages: MultiLangBench, which asks a model to build a SQLite-compatible database engine in each of 20 languages, and a COBOL migration test that requires byte-for-byte output parity after porting programs to C++, Java, Kotlin, Python or Rust. OpenAI says GPT-6 Astra averaged an 87.4% test pass rate on MultiLangBench, up from 26.8% for GPT-5.6 Sol, and GPT-6.1 Sol reached 86.6% on COBOL migration at low reasoning effort and 97.9% at higher effort, versus 5.4% for GPT-5.6 Sol. It is releasing the MultiLangBench setup and test suite on GitHub. The results are OpenAI's own.
OfficialAMD releases GAIA 0.25, its local AI agent app for Ryzen AI PCs, with a larger default model and a much smaller install
GitHub (amd/gaia) published this on Oct 18, 2026 (source: github.com).
AMD's GAIA 0.25 makes gaia chat run the flagship GAIA agent and opens the Agent UI straight to a chat that sets itself up on a fresh install. gaia init now picks Qwen3.6-35B-A3B on a 64 GB or larger Strix Halo or a 32 GB GPU and keeps Gemma-4-E4B elsewhere, with Qwen3.8 Flash Next opt-in on 128 GB Strix Halo. Dropping PyTorch, transformers and accelerate shrinks the desktop app's Python environment on Windows from about 0.9 GB to 0.1 GB, and setup now works behind TLS-inspecting corporate proxies. DeepSeek V4.1 Flash is the default cloud model on the AMD LLM Gateway and Fireworks.
OfficialAmazon Bedrock adds structured outputs for all Anthropic Claude models
AWS published this on Oct 19, 2026 (source: aws.amazon.com).
Amazon Bedrock now supports structured outputs for all Anthropic Claude models on both the bedrock-runtime and bedrock-mantle endpoints, where the model is available. Two mechanisms work on both endpoints: a JSON Schema output format that controls the shape of the response, and strict tool use that validates tool parameters against schemas. AWS says structured outputs use constrained decoding to enforce the schema during generation, so every response is valid JSON with required fields present.
OfficialAmazon Bedrock adds retrieval instructions to agentic retrieval on managed knowledge bases
Amazon (AWS) published this on Oct 20, 2026 (source: docs.aws.amazon.com).
AWS says agentic retrieval on fully managed Amazon Bedrock knowledge bases now supports retrieval instructions, which steer how the agent searches knowledge bases and how it answers. Bedrock reviews each instruction before it takes effect, versions can be pinned as immutable, and instructions are applied through the retrievalInstructionConfiguration field on AgenticRetrieveStream, per its Bedrock documentation history (October 19, 2026).
Reported by The InformationAnthropic is preparing a model that better infers what users intend from limited information, per The Information
Not confirmed by the company — reported on Oct 19, 2026 (source: theinformation.com).
Anthropic plans to release a new model this month that is better at inferring what a user wants from limited information, The Information reports. The report does not give a name, release date or pricing, and Anthropic did not comment.
OfficialSkyfall AI previews SKY-001, a world model that plans CAD and BIM designs step by step
Skyfall AI published this on Oct 19, 2026 (source: skyfall.ai).
Skyfall AI introduced SKY-001, a research-preview world model that turns a text prompt or image into a hierarchy of parts and the construction steps that build them in CAD, and turns building outlines into unit and room layouts in BIM. Skyfall says it generated a CAD design in 3.9 seconds versus 199 seconds for GPT-6 Astra, about 51x faster at roughly 1/150 of the cost, and about 119x faster on BIM; these are the company's own figures. Access is via waitlist.
OfficialLovart 2.0 adds Arti, a personal design agent with memory and scheduled design tasks
Lovart published this on Oct 19, 2026 (source: lovart.ai).
Lovart released Lovart 2.0, centered on Arti, a personal design agent that remembers a user's background, aesthetic preferences, collaboration style and frequently used assets. Users can hand Arti whole tasks through design delegation, run once or on a daily, weekly or monthly schedule with a credit limit per run. Arti also adds background design research, canvas reviews against a brief, revision plans from reviewer comments, and reusable skills from repeated workflows.
OfficialCua releases the Cua Spaces SDK to run Claude Code, Codex and other agents on computers you choose
@trycua on X published this on Oct 19, 2026 (source: x.com).
Cua introduced the Cua Spaces SDK, which lets developers put Claude Code, Codex or any agent on compute they choose and run jobs across all their computers at once. From Python (pip install cua) or TypeScript (npm install @trycua/cua), code can create a Space, run commands in it, take screenshots, watch it and delete it. Cua Spaces itself launched on Oct. 2.
OfficialTown raises $108 million Series B at a $1 billion valuation for AI assistants for teams
Town published this on Oct 19, 2026 (source: town.com).
Town, which makes AI assistants that work with whole teams, raised a $108 million Series B at a $1 billion valuation. Forerunner Ventures and Menlo Ventures co-led the round, with participation from a16z and First Round Capital, among others.
OfficialAnt Group's inclusionAI open-sources five Ming-3.0 speech models for TTS, audio understanding and ASR
InclusionAI (Hugging Face) published this on Oct 19, 2026 (source: huggingface.co).
inclusionAI open-sourced the Ming-3.0 speech family under the MIT license: Ming-3.0-tts-0.5B and tts-tiny for speech synthesis with zero-shot voice cloning, Ming-3.0-audio-flash (126B MoE, 6.5B active) and audio-tiny for audio understanding and spoken responses, and Ming-3.0-asr-tiny for Chinese, English and dialect speech recognition. On Artificial Analysis, Ming-3.0-tts scores 1,069 Elo on the controlled voice-cloning leaderboard.
OfficialCloudflare previews Clairic, a typed extraction model on Workers AI
Cloudflare published this on Oct 19, 2026 (source: blog.cloudflare.com).
Cloudflare introduced Clairic, an extraction model in preview on Workers AI that returns strongly typed values with confidence scores and probabilities. It can extract, compute, estimate and forecast values from text and up to four images, with a 65,536-token context window and $0.10 per million input tokens. Cloudflare calls it a new 'System One' class of model.
OfficialExa launches Exa 3.0 search and a cheaper Exa Lite tier at $4 per 1,000 requests
Exa published this on Oct 19, 2026 (source: exa.ai).
Exa released Exa 3.0, which it calls its largest search-quality improvement and says ranks first on every benchmark it tested, from public evaluations to coding, finance and go-to-market tasks; the results are Exa's own. It also launched Exa Lite, a search type on the same index as Exa Auto at $4 per 1,000 requests, versus $7 for Auto.
Officialfal launches FLUX 3 Max, a post-trained FLUX 3 video model it says makes a 10-second clip in about 12 seconds
Fal published this on Oct 19, 2026 (source: fal.ai).
fal released FLUX 3 Max, a video model post-trained by fal Research with Black Forest Labs for text-, image- and keyframes-to-video. fal says it generates a 10-second video in 12.45 seconds on average, by its own benchmark. In Artificial Analysis's initial image-to-video leaderboard it scored about 1,139 Elo, tied with Seedance 2.5.
OfficialNotion rebuilds its AI agents with their own computer and day-long tasks, and moves to a new credit system
Notion published this on Oct 19, 2026 (source: notion.com).
Notion says Notion Agent and Custom Agents now run on new infrastructure with their own computer and browser and can work on a single task for more than 24 hours, up from five minutes. Its upgraded Auto model router picks a model per task, which Notion says matches frontier models at about 90% lower average cost. Business and Enterprise seats now get monthly Standard and Select credits; the credit system applies now to new workspaces, and existing workspaces move over on Nov. 2.
OfficialGitHub deprecates six models in Copilot, including GPT-5.5, GPT-5.4 and Grok 4.5, as of Oct. 19
GitHub published this on Oct 19, 2026 (source: github.blog).
As of Oct. 19, GitHub has deprecated Gemini 3.7 Flash, GPT-5.5, GPT-5.4, GPT-5.4 mini, GPT-5 mini and Grok 4.5 across most GitHub Copilot experiences, including Copilot Chat, inline edits, ask and agent modes, and code completions. GitHub suggests Gemini 3.8 Flash, GPT-6.1 Sol, GPT-6 Luna and Grok 4.7 as replacements; the deprecations were announced on Sept. 18. Copilot Enterprise administrators may need to enable the alternative models through their model policies.
OfficialMicrosoft Foundry adds Meta's Muse Spark Open, NVIDIA's Nemotron 3.5 Super VL and Reflection's Beam 1.0
Microsoft published this on Oct 19, 2026 (source: techcommunity.microsoft.com).
Microsoft says three open-weight Hugging Face models are now available in Microsoft Foundry: Meta's Muse Spark Open (816B total and 42B active parameters, in NVIDIA's 4-bit NVFP4 format), NVIDIA's Nemotron 3.5 Super VL (120B total, 12B active) and Reflection's Beam 1.0 (501B total, 23B active). Muse Spark Open runs with a 1,048,576-token context on eight NVIDIA H200 GPUs per instance, Nemotron 3.5 Super VL gets 524,288-token presets on four A100 or H100 GPUs, and Beam comes in BF16, FP8 and NVFP4 versions with up to 524,288 tokens. Microsoft cites publisher-reported scores, including 81.6% on SWE-bench Verified for Beam and 89.2 on MCP-Atlas for Muse Spark Open.
OfficialNace AI releases Drex Max, a multimodal decision model that picks computer-use agent actions from screenshots
Nace.AI published this on Oct 19, 2026 (source: nace.ai).
Nace AI released Drex Max, a decision model that reads text and images and returns a calibrated probability for each answer option instead of generating text, aimed at agents that operate computers and browsers. It is available only through Nace's hosted API as drex-max, at $0.06 per million input tokens with output tokens not billed, and accepts up to eight images per request. Nace says Drex Max scores 67.59 on Decision Index 0.3.2, 6.9 points above Jev 1.14; the scores are Nace's own.
Statement by Zohran Mamdani (New York City)New York City Mayor Zohran Mamdani says the federal government has abdicated on AI and urges New Yorkers to push back, per TechCrunch
Zohran Mamdani (New York City) said this on Oct 19, 2026; not an official company announcement (source: techcrunch.com).
Speaking at the AI Now Institute's People's AI Assembly in Manhattan on Monday, New York City Mayor Zohran Mamdani urged New Yorkers to reject the idea that AI's rapid rollout is inevitable and said public officials should hold AI companies accountable. He criticized what he called a "complete abdication of federal law" on AI and said the national debate is too often framed as an arms race with China. TechCrunch notes that Mamdani ordered a ban on AI use in elementary and middle schools last month and introduced a policy requiring disclosure when property listing photos or videos are made with AI.
OfficialCognition makes Devin Oncall available to all organizations
Cognition published this on Oct 19, 2026 (source: docs.devin.ai).
Cognition first launched Devin Oncall on Oct. 15, and its Oct. 19 release notes now make it available to all organizations. Oncall responders watch Slack or Microsoft Teams channels and PagerDuty incidents, triage new alerts, bug reports and support requests, and investigate them with an organization's telemetry tools, code, deployments and runbooks to post a likely root cause and recommended next actions. The release notes also add a rules editor for Devin Review, a one-off "Review now" option for unreviewed commits, and CI status checks for commits, branches and tags on GitHub, GitLab, Azure DevOps and Bitbucket.
Reported by ReutersOpenAI is racing to raise about $30 billion at a valuation of over $1.4 trillion before Anthropic's IPO, per Reuters
Not confirmed by the company — reported on Oct 19, 2026 (source: money.usnews.com).
OpenAI is seeking investor commitments for a round of about $30 billion that could value it at over $1.4 trillion, aiming to finish before Anthropic's expected public market debut in mid-November, according to two people familiar with the matter. SoftBank Group, Abu Dhabi-based MGX and other existing backers are in talks to join the round, one of the people said. OpenAI is pitching investors on $50 billion in annualized revenue, which it expects to reach $70 billion by year-end, according to one of the people. OpenAI was last valued at $852 billion in March.
OfficialNVIDIA opens some Barcelona Supercomputing Center AI Factory services to Inception startups through its portal
NVIDIA Blog published this on Oct 19, 2026 (source: blogs.nvidia.com).
NVIDIA says startups in its Inception program can now reach some Barcelona Supercomputing Center (BSC) AI Factory services through the Inception portal, as BSC's AI Factory Accelerator Programme moves into full operation after a pilot. The programme gives startups sovereign compute, technical support and training, and NVIDIA says the BSC AI Factory, Spain's first, has delivered more than 800 services since its April 2025 launch. NVIDIA names three Inception members, water-network platform Qatium, ECG spin-off OneCareAI and pharma agent maker Avatia, as the pilot's AI Champions, and says an upcoming MareNostrum 5 extension will add supercomputing access through European AI factories.
Daily AI news, sourced