Daily AI News

OpenAI has published a National Community Compact alongside a Compute Campus Policy Blueprint, describing them as a framework for responsible AI infrastructure development. According to OpenAI, the two documents are meant to protect local resources, create jobs, and guide how AI data centers and related facilities are built. The company did not detail specific commitments or timelines in its summary of the announcement.

OpenAI has announced it will acquire Pierre Computer Company, Inc. The company said the deal expands Codex's tools for building, understanding, and reviewing code. The announcement names Codex as the product affected and does not state the purchase price or a closing date. Further details on timing and availability were not provided in the announcement.

OpenAI says it built an autonomous pipeline that found and validated 948 local privilege-escalation vulnerabilities in the Linux kernel, then generated patches for them. The company published a short summary of the work. The summary does not name the models behind the pipeline or describe how the patches are being routed to Linux maintainers.

OpenAI says Daybreak Red is now available in AWS GovCloud (US-East), giving approved customers access to GPT-5.6 Cyber through Amazon Bedrock. The offering is positioned for authorized cybersecurity work, and availability is limited to customers who have already been approved for it in that region.

Moonshot AI has published the weights for Kimi K3.1 Pro, an open-weight model that handles both text and images and the first release in its K3.1 series. It is a mixture-of-experts design with 2.8 trillion total parameters and 104 billion active per token, and it accepts context windows of roughly one million tokens. The company says it post-trained the model from Kimi K3 using large-scale joint reinforcement learning across hundreds of task domains, and it reports improvements in coding, 3D creation, financial research, factual accuracy and other agentic work. Weights are downloadable on Hugging Face in 8-bit precision, and an API is also offered. The model card includes benchmark comparisons against other open and proprietary models.

Ricursive Intelligence has released Ripple-65B Preview, a general-purpose open-weight language model derived from the internal system that powers its AI chip design platform. According to the company, the model was pretrained on 12 trillion tokens, handles context up to 1M tokens, and uses a dense Transformer architecture rather than a mixture-of-experts design. Ricursive reports a TerminalBench 2.1 agentic coding score of 80.52, which it describes as U.S. open-weight state of the art, ahead of several far larger open-weight releases. Ripple-65B Preview is downloadable on Hugging Face and served through Together AI, GMI Cloud and OpenRouter, with Fireworks going live later. Ricursive puts its total research and development cost at $32 million.

Anthropic published research on provable inference, in which a model provider commits to a specific model and attaches a cryptographic proof that each response came from it. The work centers on FlashProve, a zkSNARK-based prototype. Anthropic released five artifacts: the blog post, a technical report, a Claude-written protocol specification, executable code with a prover and two verifiers, and Lean proofs with an interactive visualizer.

Gradium published a closer look at Phonon, its small text-to-speech model that runs on a local device with no server or network connection. Two checkpoints of roughly 65 MB and 75 MB cover English, French, German, Spanish and Portuguese, with more than 100 built-in voices plus voice cloning and text-described voices in Gradium Studio. The English 31m model is Apache 2.0 and usable in commercial products; the multilingual 40m model is CC BY-NC for non-commercial work. Phonon ships in Transformers from v5.21 and through ptts, a new inference framework with CPU, GPU and NPU backends. On Hugging Face's Open TTS Leaderboard it placed first for German and Spanish intelligibility and in the top three for every language the company evaluated.

Google Research introduced Agile Deliberation, a framework for building image classifiers around subjective categories such as healthy food or content-moderation cases. It guides a user through concept scoping, where broad ideas are split into subcategories and exclusions, then concept iteration, where a vision-language model surfaces borderline images and routes feedback into an automatic prompt-refinement loop. The team open-sourced the code and algorithms on GitHub for the research community. In a live study with 18 sessions, participants using the workflow produced classifiers that outperformed zero-shot, automated collaborator, and manual deliberation baselines, the researchers report.

Dharma-AI has released the Dharma Field Extractor, a 2-billion-parameter model that takes a document image and returns structured, typed JSON instead of raw text. Two request modes are supported: an open mode that identifies fields from layout and context, and a schema-directed mode that returns exactly the key names and types supplied, using null where a requested field is absent. On the team's 546-document benchmark it scored 0.834 field-level F1, above the four frontier configurations they tested, while running self-hosted on a single L40S.

A preprint on arXiv presents NEO-VLA, a vision-language-action framework that drops the separate visual encoder used in most manipulation policies. Images are split into patches by lightweight convolutional layers and then processed alongside language through the same backbone, so a flow-matching action expert can draw on instruction-conditioned visual features at several depths. In controlled comparisons, the authors report that this native design beats encoder-based versions, that full fine-tuning narrows the gap without closing it, and that added scale alone does not recover it. Layer-wise probing found action-critical signals often peak well before the final visual representation.

NVIDIA published the AI Factory Operations Agent it uses internally to run its own DGX SuperPOD clusters, and released the framework and code on GitHub along with a developer example others can deploy. A central agent routes site reliability engineers' natural-language questions to specialized sub-agents covering job recovery, hardware faults, observability, Kubernetes state, cluster management and documentation search, and returns findings through a chat interface. The stack is built on NVIDIA's NemoClaw blueprint and OpenShell runtime and can run entirely on-premises with locally served Nemotron open models. NVIDIA says it runs across its DGX H100, B200 and GB300 production clusters, with sandboxed sessions and read-only cluster access.

LangChain published version 1.7.2 of its Fireworks partner package, the library that connects LangChain applications to Fireworks models. The update reports streamed token usage a single time, adds the served model name to response metadata, and moves reranking support out of beta. It also refreshes the bundled model profile data, relaxes upper bounds on third-party dependencies and bumps packages including certifi, aiohttp and orjson, and stabilizes a test for forced tool-call extraction. The release is available now to Python developers using the langchain-fireworks package, following the 1.7.1 version.

NVIDIA published details of CUDA Toolkit 13.5, which moves support for its Vera Rubin GPU architecture from early developer access to general availability for production use. Windows on Arm support, general since CUDA 13.4, now extends to DGX Station for Windows. Scheduling changes include locality domains that can include all device SMs, and Multi-Process Service v3 adds weighted timeslicing, SM partition borrowing and lending, cgroup-based routing and per-namespace partition assignments. CUDA Tile gains multi-GPU communication over NVLink and NVSwitch plus a preview of SIMT interoperability, while CUDA Python's cuda.core 1.2 adds batched memory copies. The toolkit is available to download now.

The manager agent runs on a model such as NVIDIA Nemotron, uses the NemoClaw harness for tools, persistent memory and delegation to specialist agents, and relies on NVIDIA OpenShell to isolate execution and enforce access policies that determine which decisions need operator review. Specialists cover visual inspection via NVIDIA Cosmos, fleet management with Isaac Sim and ROS 2, and root cause analysis using Kumo Anomaly. A sample SMT line follows an AOI alert through secondary inspection, robot rework and a root cause report. FOX is offered through early access and can run fully on-premises on DGX Station.

AWS introduced Smart Approvals for Amazon Quick, a feature that reviews an AI agent's proposed action against the user's request and either approves it automatically or sends it back to the user to decide. The company says it checks each action against Custom Permissions, Quick Guardrails, connector action permissions and user settings first, so denied actions stay denied and mandatory human approvals still apply.

The release adds standardized refusal content blocks, so a model declining a request is represented in one shared format, and it translates Gemini safety blocks into those refusal blocks. It also ends a tool run when a request is cancelled and fixes loading of RemoveMessage objects that were serialized by LangChain.js.

Cloudflare released new analytics and access controls that let site owners see which personal AI agents visit their sites and decide how much access each one gets. Sites using Account Abuse Protection, in early access, can see which user an agent represents. In Security Settings, the AI Crawler and Agent access policies setting, renamed from Agents policy, now offers per-agent choices: allow, block on pages with ads, or block, with the same options offered during zone onboarding. Existing Search and Training policies are unchanged.

The llama.cpp project released version 0.9.0 of its open-source local inference engine. The update adds an OCP FP8 E4M3 tensor type in saturating mode, with conversion helpers and a new ftype, plus a flash_attn_ext_rows op that makes the unified KV cache cheaper for multi-sequence inference. A batched tensor copy API replaces the older _2d calls, implemented for Metal and CUDA. Four architectures are now supported: Clover-35B-A3B-Preview in autoregressive mode, Limite 1B Violetto, Metis-1.6, and Mistral Large 4. Gemma 4 E2B/E4B get prefill optimizations, and ggml v0.29.0 brings NCCL multi-node tensor parallelism.

Vercel says Kimi K3.1 Pro and its faster variant, Kimi K3.1 Pro Highspeed, can now be reached through AI Gateway from additional providers that hold Zero Data Retention agreements. Moonshot AI had been the only source for both models until this update. Both take image input and support a 1M-token context window; the Highspeed version is aimed at interactive coding agents. Developers call them as moonshotai/kimi-k3.1-pro or moonshotai/kimi-k3.1-pro-highspeed through the AI SDK, Chat Completions, Responses, or Anthropic Messages APIs, or from a coding agent such as Claude Code, Cursor, or Codex after installing the latest Vercel CLI and running setup.

Available now are connections to Google Workspace and Microsoft 365, Conversation History, reusable Skills, and the per-user option. General availability in November 2026 covers Coworker in Microsoft Teams, iOS, Mac and Windows apps, a Coworker MCP server, and a free trial for Flex Credits customers. Daily Digest, scheduled tasks and Coworker in the Field Service, Consumer Goods Cloud and Life Sciences Cloud mobile apps are expected in December 2026. Salesforce customers can turn it on in Setup; The Adecco Group is expanding it across 40 countries and Fulton Bank runs more than 20 use cases.

OpenAI has introduced Images 2.6, an update to its image generation model. According to the company, the release adds faster image generation, GIF creation and more precise editing, along with new controls in the API. The update is available across ChatGPT, Codex and the API, so both chat users and developers building on OpenAI's platform get the same new capabilities.

Google DeepMind introduced a suite of Cheminformatics Skills for Google Antigravity, turning the agent surface into a workbench for drug discovery tasks. The skills pair a language model that handles workflow orchestration and scientific reasoning with deterministic chemistry tools built on RDKit, plus connectors to ChEMBL and PubChem, so an agent can extract molecules from literature and catalogs and reconcile different structure formats.

Ai2, the Allen Institute for AI, released StateBench, a test set for measuring whether language models can follow changes to information about people, objects and variables across long documents. The benchmark has four difficulty levels, from values that never change to updates that depend on earlier updates, and presents the same problems in several formats. Ai2 published a technical report, data and code. In its evaluations, several open pretrained models and four frontier models handled easier levels but lost accuracy on harder ones as inputs grew; the longest general finite cases fell to 13% or lower. Models trained directly on the tasks showed recurrent and hybrid designs generalizing farther than transformers.

IBM Research published a technical article on Palette, a 20-billion-parameter open-weight model fine-tuned to build slide decks that follow a given design language, demonstrated with the IBM Design Language. The work splits the task between two roles handled by a single LoRA adapter: a designer that writes a structured JSON brief covering palette, type and each slide's layout and content, and a coder that turns one slide's spec into pptxgenjs JavaScript. A harness renders each slide, flags overlap or overflow, and keeps an edit only when the checker confirms improvement. On PresentBench, Palette placed first in two of three domains measured.

A paper on arXiv describes TurnWise, a reinforcement learning framework for streaming video understanding, where a model repeatedly decides whether to answer or keep observing. The authors note that existing GRPO-based approaches keep a single trajectory-level advantage across turns, which they say can reinforce redundant replies after earlier successes and discourage useful follow-ups after failures. TurnWise instead assigns credit per turn, using silence-completed prefix scoring to build group-relative advantages from each decision's remaining return. It also averages token losses within each turn before averaging across turns, so responding and waiting carry equal weight regardless of output length. The authors report it outperforms prior RL approaches in their experiments without a learned critic or extra rollouts.

Meta detailed F3 Lightning, an infrastructure layer built on top of its Facebook Feature Framework (F3), which compiles and serves the machine learning features behind ads ranking models. F3 Lightning separates the release boundary from the execution boundary, so each feature is compiled, validated, deployed and rolled back on its own, while a shared execution plan assembled at startup keeps the compute savings. Meta says comparable extraction features now reach production about 3.6x faster with no increase in infrastructure compute cost.

AWS published a post detailing how Cartesia, a partner focused on real-time voice AI, runs its Sonic text-to-speech and Ink speech-to-text models on AWS infrastructure. Sonic is offered as an Amazon SageMaker AI model package through AWS Marketplace, so teams subscribe once and host the model on an endpoint inside their own virtual private cloud, keeping inference data in their account. Ink is reachable through the Cartesia API, purchased via the Cartesia Voice AI Platform listing in AWS Marketplace.

Amazon Web Services now offers in-country inference in India for Kimi K3, the open-weight, multimodal foundation model from Moonshot AI, on Amazon Bedrock. AWS says traffic between Regions stays encrypted on its own network. The model reads images and PDFs directly, so scanned documents can be queried without a separate OCR or translation service. The India option is available now to Bedrock customers, subject to the required IAM permissions and Region settings.

Databricks says its MLflow platform is now a supported destination for native OpenTelemetry export from the Claude Platform, released in Public Beta. Traces from Anthropic's SDKs, the Messages API and Claude Managed Agents arrive through Zerobus Ingest and are stored in Unity Catalog tables tied to an MLflow experiment, where they can be inspected alongside traces from other services or queried with SQL. Setup takes three steps: create an experiment with a Unity Catalog storage location, configure the Databricks destination using workspace ID, region and spans table, then run an agent and open the trace. Captured inputs and outputs require opting into the "full" content mode. MLflow Issues is also in Beta.

OpenAI is raising the rewards it pays through its Security Bug Bounty program and adding new bonus categories for independent security researchers. The added bonuses cover report quality, loyalty, and priority campaigns. The changes apply to outside researchers who report security issues to OpenAI; the announcement does not state the size of the new payouts or when they begin.

OpenAI introduced a Grid Protectors Program, which the company says will involve working alongside utilities. The stated aim is to help future compute campuses support grid reliability and to protect consumer electricity costs. The announcement offers no details on which utilities are participating, what the program's commitments are, or when work begins.

OpenAI has introduced LifeVal, a benchmark that measures how well AI helps people complete everyday tasks. The company cites examples including finding childcare, managing finances and handling government paperwork. LifeVal is presented as a way to evaluate AI performance on the ordinary, real-world work people bring to these tools.

14, an update to its model runner that changes how its launch commands and API endpoints handle model settings. The `ollama launch claude` command now applies each model's output token limit for Claude Code, so cloud models no longer hit the 128k default cap. The release also fixes decoding of cloud web search responses on /api/chat, preserves readable input content in /v1/responses including prior assistant refusals, lets the Ollama app retry ChatGPT disconnect failures, has `ollama launch codex` verify ChatGPT routing before configuring, and speeds up top-k sampling on MLX for large vocabularies.

Collinear AI has published CWE-bench v2, a public benchmark of 100 security tasks covering 91 different weakness classes, announced on the Hugging Face blog. Each task gives an agent a checkout of a real open-source repository and a short instruction to find and fix a vulnerability without breaking existing functionality, and all scoring is programmatic, so results are deterministic and reproducible at no extra cost. A held-out set is available by request. The public set, grader and harness are open, with every task shipped as a Dockerized Harbor environment and tagged as an RL environment that can also reward training defensive agents. On the project's leaderboards, Gemini 4 Argon leads with 67.8% pass@1 and 81.0% pass@4, just ahead of Claude Opus 5.5, and 95 tasks are solved by at least one of the 11 evaluated models.

Hugging Face published transformers-mic, a proposed standard for packaging lightweight regression tests for models integrated into its Transformers library. Each contract combines a fixture shrunk from a real checkpoint's configuration, a runnable example of a common usage path, and human-reviewed baselines, then checks on a cheap CPU runner that the example still produces those results. The post says a small UMT5 fixture would have caught an attention bug that sat in main from July 30 to September 28, 2026. A command-line tool, tmic, builds contracts locally with review at each step, and published contracts live in the huggingface-mic dataset. The contract job in Transformers CI currently reports failures without blocking pull requests.

A group of researchers has proposed LoCo, a training method that teaches reasoning models to conclude once they already have enough information instead of continuing to deliberate. The paper, posted on arXiv, identifies candidate overthinking states inside a model's own traces, freezes the reasoning up to that point, and regenerates shorter continuations that are verified. State-conditioned preference pairs from those pairs are then used for direct preference optimization, so the model learns when to keep going and when to stop. All training data came from self-generated mathematical traces. Completions shrank 41.7%–45.4% on five math benchmarks and 23.4%–36.3% on three code benchmarks, with accuracy maintained or improved. The authors say code and model will be released on GitHub.

The authors of CoworkOS describe a general-purpose harness for long-horizon productivity work that splits task strategy from execution management. The model chooses actions, delegation and acceptance, while the runtime holds state and feedback. A three-layer design lets work run directly or be handed to Subagents, synchronously or asynchronously, and a shared application layer lets people and agents inspect and revise the same documents, presentations, CAD models and Blender scenes. Hierarchical workspaces and persistent semantic memory keep files, decisions and job logs resumable after an interruption.

They report that ideal per-input routing beats the best single fixed candidate by 5.4 points, and that learned routers beat it by 0.3 points across 16 benchmarks, 2.1 points on video generation and 1.1 points on robot manipulation while using 24% less policy time. Routing raised efficiency 8.82x on average over the reference model. The paper is available on arXiv.

Researchers introduce DreamX, a method that turns a pretrained video diffusion model into a learner of compact latent actions. The resulting latent space is presented as a shared action interface across human and robot embodiments, environments, and action parameterizations. DreamX reports state-of-the-art results among latent action models on LARY-Bench and supports any-to-any video transfer, zero-shot action transfer, and policy learning with lower computational cost. The work is described in an arXiv paper; no code or model release is stated in the abstract.

A team of researchers has published CUA-Master, a training framework for computer-use agents that pairs an outer loop, which keeps pulling in open-world material and targets whatever the model currently handles poorly, with an inner loop that turns real workflows mined from instructional videos into reusable building blocks for generating new tasks. Trajectories are gathered in live environments using a retry scheme that rolls back to earlier snapshots when a run fails. The work is a preprint on arXiv; no code or checkpoint release is described in the abstract.

The researchers behind Kling have published a technical report on arXiv arguing that video generation quality now hinges more on semantic control than on rendering, and walking through the understanding stack they built around that idea. The report introduces Kling-MLLM, a shared multimodal foundation model that handles image, video, audio, synchronized audio-video and interleaved inputs, with specializations for each downstream module. It covers captioning treated as a dense semantic contract, prompt enhancement refined with reinforcement learning, paired reward models for critique and fast scoring, and automatic evaluation protocols designed to expose scoring shortcuts. The paper is on arXiv; no code or model weights are announced in the abstract.

0 is out, a release the project says spans 927 commits from 389 contributors.

NVIDIA has released NVIDIA Attribution as open source, a system that ties failures in distributed AI jobs back to their actual origin instead of the point where training stopped. It ships as two pieces: a Service that exposes a REST API around the attribution engine, and a Blueprint, a Python orchestrator that pulls scheduler and telemetry data from Slurm, Kueue, Loki, Kubernetes pod logs, local files and PromQL metrics. NVIDIA says it has run the system in production across hundreds of thousands of GPUs, analyzing millions of workloads weekly. The repository, a quickstart with sample workloads and a local diagnosis dashboard, and Blueprint configuration docs are available now for teams running their own clusters.

NVIDIA published NVIDIA Recovery, an open-source, workload-level service that watches a cluster's training jobs and automates a four-stage recovery loop: detecting crashes, hangs and stragglers from application logs, GPU metrics and scheduler signals; attributing the failure; isolating the implicated nodes; and then restarting or stopping workloads according to operator policy. It sits in the company's DSX OS software stack and includes a Slurm integration through a reference Advisor Controller installed via Helm chart in either in-cluster or SSH mode. Source code and an issue tracker are in the dsx-ai-factory/NVIDIA-Recovery repository, alongside a quickstart that runs locally without a physical GPU and an operator deployment guide. New deployments begin in monitor mode, publishing recommendations without taking scheduler actions.

NVIDIA has published a preview recipe family in its AI Cluster Runtime (AICR), part of the DSX AI Factory Platform, that records the settings needed to stand up Kubernetes clusters on Vera Rubin NVL72 racks. 0, and come in training, inference and Dynamo inference variants. NVIDIA says the bundles passed deployment and conformance checks plus a cross-node NCCL all-reduce performance gate.

NVIDIA has released new Cosmos 3 Nano simulation recipes and example checkpoints, along with an expanded Cosmos Framework for training, optimizing, and running world models for physical AI. The framework now covers models including Cosmos 3, Alpamayo, GR00T, Qwen, and pi0.5, and offers agent skills and recipes for autonomous-vehicle, robotics, and surgical simulation. The released checkpoints generate synchronized camera video and LiDAR observations and support PyTorch and vLLM-Omni on Hopper and Blackwell GPUs. NVIDIA also describes Cosmos NIM microservices for accelerated inference, with configuration guidance for GPU memory and deployment. The tools are aimed at physical AI developers who need to adapt world models to their own sensors, tasks, and data.

NVIDIA detailed Workload Map, a project that gives Kubernetes workloads a shared structural description so schedulers, controllers and platforms can read distributed jobs without framework-specific parsing code. It ships as a Kubernetes CRD plus a Go library, along with a catalog of prebuilt definitions covering JobSet, PyTorchJob, RayCluster, RayJob, RayService, KServe, Knative, MPIJob, NVIDIA NIM, LeaderWorkerSet, Milvus, Grove PodCliqueSet and NVIDIA Dynamo. The project, formerly called Karta, is available now from NVIDIA's repository, and its CLI installs through Homebrew. KAI Scheduler and NVIDIA Run:ai already consume the definitions, and Elotl's Nova uses them for multi-cluster scheduling. Teams can add other workload types by writing their own YAML definition.

LangChain published version 1.4.8 of its Python package, a maintenance release that bumps langchain-core and refreshes a batch of third-party dependencies, including fastmcp, opentelemetry-api, authlib, pyasn1 and joserfc. Alongside the changelog, the maintainers said Python 3.10 support will be dropped in langchain-core 1.8.0 and langchain 1.5.0, making the current minor versions the last ones that run on 3.10; users are asked to upgrade to Python 3.11 or later. The release is available now from the project's GitHub repository.

LangChain published langchain-core 1.7.4 on GitHub. Compared with 1.7.3, the patch includes the release commit and a fix in the Anthropic translator that stops message content from being mutated. The release notes also carry a deprecation warning: Python 3.10 will no longer be supported starting with langchain-core 1.8.0 and langchain 1.5.0, and the current minor versions are the last ones that work on it.

Vercel added Tencent's Hy4 to AI Gateway, listed at 50% off until December 31, 2026. The model handles reasoning, tool calling, structured outputs and image input, and carries a 1M-token context window for working across a codebase or large document sets. It is the first Hy model with vision, which Vercel says lets it read complex charts and pick out fine image detail, and it supports computer use in both command-line and graphical interfaces. Compared with Hy4 Preview, it follows instructions more closely and performs better on long context, long-running tasks and deep research. Developers call it as tencent/hy4 through AI SDK or Chat Completions, or select it in coding agents such as Claude Code, Codex and Cursor after installing the latest Vercel CLI and running setup. A model playground is also available.

Salesforce announced it has signed a definitive agreement to acquire Cresta, whose voice-first platform runs AI agents that handle customer service calls end to end and also guides human representatives in real time. Cresta's agents work on complex, high-volume calls for brands including Marriott, Comcast and Alaska Airlines. The transaction is expected to close in the first quarter of fiscal 2028, subject to customary closing conditions and required regulatory approvals. Salesforce said Cresta will continue to support the contact center and telephony platforms its customers rely on today.

OpenAI says Qodo has built a plugin that pulls company knowledge and coding standards into Codex for AI-assisted code review. The goal is to help developers spot P0 issues across their repositories before they open a pull request.