AI news for Nov 18, 2026
28 stories
- ResearchOpenAIOpenAI examines how coding agents are speeding up AI research
OpenAI has published an analysis of how coding agents are affecting the pace of AI research. The company says it measured where these agents accelerate research work and where their performance remains uneven, and it discusses what that pattern suggests about recursive self-improvement. The analysis is available on OpenAI's website.
- ModelsGoogle DeepMindGoogle DeepMind releases Genie 4 world model in Project Genie
Google DeepMind and Google Labs are rolling out a rebuilt Project Genie experience that gives users first access to Genie 4, which the teams describe as their most realistic and interactive world model so far. Genie 4 lets people type a new prompt mid-session to change weather, add characters or redirect action, and it adds real-time native audio such as ambient sound and character dialogue.
- ProductsGoogleGoogle brings the Gemini app to lower-memory Android Go phones
Google said the Gemini app now runs on eligible Android (Go edition) phones, after the company rebuilt the app so its core experience works on hardware with less memory. The post is by Angus Logan of the Gemini App team. Users on those devices can speak prompts instead of typing them, hold spoken interview practice in more than 45 languages through Gemini Live, aim the camera at notes or forms to learn and translate, and generate images with Nano Banana. The app can be downloaded from Google Play, opened by pressing and holding the power button on supported phones, or found preloaded on upcoming devices from partners such as Transsion.
- ResearchHacker NewsEnergy-Based Reasoners hit 99.9% on Sudoku with 345x less inference compute
Researchers at Duke, UIUC and Harvard introduced Energy-Based Reasoners, a class of reasoning models that learn an energy function over latent reasoning states and treat looped updates as gradient descent on that energy. The energy score lets the model choose among candidate answers without an external verifier, while its gradient indicates how to refine a state and when to stop looping. With 1.8M parameters, the method reached 99.9% accuracy on Sudoku-Extreme using 345× less inference compute than the next-best model, and state-of-the-art coverage on multi-solution N-Queens and Graph Coloring benchmarks. The work is presented in a paper and project page; no code release or deployment is mentioned.
- ModelsHacker NewsGoogle releases Gemini 4 Neon, a faster lower-cost frontier model
Google DeepMind announced Gemini 4 Neon, which it describes as delivering frontier-level intelligence with fast end-to-end task completion across software engineering, knowledge work and science workflows. The model has a 1 million-token context window and is priced at $3 per million input tokens and $15 per million output tokens, with cached input at $0.15 per million. Google reports a score of 61 on the Artificial Analysis Intelligence Index at its xhigh thinking level and a second-place ranking on the Vals Index, and says it records the company's lowest fabrication rate yet in internal Antigravity testing. Neon is available today in the Gemini API and Google AI Studio, in Gemini for business, and for paid subscribers in Google Antigravity and the Gemini app. A cyber-focused variant is rolling out to trusted defenders through the Fairwind Program.
- ModelsHacker NewsMicrosoft AI launches seven MAI models and Maia 200 chip at Ignite
Microsoft AI launched a family of seven models at Ignite 2026, including MAI Code 2, which the company says reaches 98% of GPT-6.1 Sol's coding quality at about 40% of the price, and MAI Code 2 Flash for high-volume routine work. Code 2 is available now in GitHub Copilot, and Code 2 Flash is in Copilot plus a limited preview in Microsoft Foundry.
- DevelopersGemini CLIGemini CLI v0.69.0 adds token usage heatmap and custom theme files
The update adds a /stats heatmap subcommand that charts daily token usage, and support for a standalone theme.json file so users can define custom themes without touching other configuration. The /model command now validates model names and accepts direct arguments. Users can also override the 60-second API headers timeout through a setting (model.requestTimeoutMs) or an environment variable. The release carries a long list of fixes covering macOS sandboxing, @file references, piped stdin, retry handling and memory retention. It is available now on GitHub for developers using the CLI.
- DevelopersAnthropic Claude CodeAnthropic turns on auto permission mode for Claude Code on cloud providers
Anthropic shipped Claude Code v2.1.328, which enables an auto permission mode for Claude Opus 4.6 and Sonnet 4.6 when the tool runs on Amazon Bedrock, Google Cloud's Agent Platform, Microsoft Foundry or the Claude apps gateway. Interactive sessions started without a permission mode now begin in it.
- Open sourceHugging FaceRadixArk releases DaisyRose-27B, an open agent model built on Qwen3.8-27B
RadixArk has released DaisyRose-27B, its first model, post-trained from Qwen3.8-27B with reinforcement learning on long-horizon agent tasks. Checkpoints are published under Apache 2.0 in BF16, FP8 and NVFP4, and the model keeps the base model's 256K-token context window and image understanding. On Terminal-Bench 4.1, pass@1 rose from 5.6% to 12.0%, and RadixArk reports smaller gains on other coding, reasoning and instruction-following benchmarks. The team recommends serving it with SGLang.
- ProductsHugging FaceIBM time-series models now run natively inside Confluent
IBM Research published a walkthrough of its time-series foundation models running natively inside Confluent's data streaming platform. The family of four models, which IBM says has passed 50 million downloads, is available as built-in functions for Apache Flink: fully managed in Confluent Cloud with nothing to deploy, or on premises through Confluent Platform. The models run on CPUs without GPUs. IBM describes them as handling forecasting, anomaly detection and optimization on streaming data, applied through SQL, with example scenarios in manufacturing, retail, banking and airlines. Teams can start from a pretrained model, fine-tune it, or combine its output with other business logic.
- ResearchHugging Face PapersResearchers Release Mind2Act World Benchmark for Robotic Manipulation
The paper's authors introduce Mind2Act World, a diagnostic benchmark built from 30 simulated manipulation tasks drawn from five real-world application domains. Each task carries three difficulty levels and two scene settings, Clean and Random, giving 180 configurations. The tasks are split into three views: Mind, which tests decisions based on past observations; Act, which tests precise and timely execution; and Mind2Act, which combines both in closed-loop tasks where action outcomes inform later decisions. A diagnosis harness attributes failed episodes of three representative systems to reasoning, execution or composition, and summarizes each in a profile card. The authors evaluated multimodal agents, vision-language-action models and world-action models.
- Open sourceHugging Face PapersWeMM-Video releases open-source models for audio-visual video understanding
The WeMM-Video report introduces a family of Chinese-English bilingual omni-modal models at 2B, 4B, 9B and 35B-A3B sizes, built around joint audio-visual perception and reasoning. They handle text, images, audio and video in one system, with an instruct mode for direct answers and a thinking mode for step-by-step reasoning. Training covered roughly 1.2 trillion tokens using progressive pre-training plus multi-stage post-training with multi-teacher OPD, anti-repetition MrDPO and specialist reinforcement learning. The authors report evaluations across more than 120 benchmarks, saying the flagship 35B-A3B model sets a new open-source state of the art on omni-modal video tasks, beating Qwen3.5-Omni-Flash and Seed2.0-Mini-0428 and narrowing the gap to Gemini-3.0-Flash, while the 9B model surpasses Qwen3-Omni-30B-A3B-Instruct. The model family, evaluation and reproduction pipeline are publicly released for the research community.
- ResearchHugging Face PapersCoda-DaVinci framework lets agents steer video generation with 3D proxies
A new arXiv paper studies Code2Video, an approach where an agent builds 3D proxies that guide a video generator, giving explicit control over scene structure and motion. The authors ran controlled experiments and a blinded human evaluation on proxy granularity, physics simulation and prompting. They report that coarse proxies favor natural motion while medium-grained ones favor consistency, and that physics simulation plus proxy-aligned prompts raise average quality. From those findings they built Coda-DaVinci, a task-adaptive agentic framework that picks between proxy-guided and direct generation, sets modeling granularity per entity, and uses a multi-agent harness to keep scene assets and guide repairs through visual inspection and independent review. The work is published as a paper; the abstract describes no code or demo release.
- ProductsNVIDIA NewsroomNVIDIA ships RTX Spark compact desktops for local AI work
NVIDIA says RTX Spark compact desktops from ASUS, Dell and MSI are available now, joining the RTX Spark laptops already on sale. The systems pair a 20-core Grace CPU with a Blackwell RTX GPU and up to 128GB of unified memory, and NVIDIA says they can run 120-billion-parameter models locally. Acer, GIGABYTE and HP mini PCs are listed as coming soon, along with a Microsoft Surface RTX Spark Dev Box.
- ProductsAWS Machine LearningAWS adds full fine-tuning to SageMaker AI serverless model customization
Amazon SageMaker AI Serverless Model Customization now supports full fine-tuning, AWS said, giving teams an alternative to parameter-efficient methods such as LoRA without provisioning GPUs, setting up clusters or configuring distributed training. Users supply a base model, dataset and hyperparameters, and the service handles compute allocation and checkpoint management. Switching between the two training modes requires one setting change in the SageMaker Python SDK.
- ProductsAWS Machine LearningAWS adds workload recovery and new health checks to SageMaker HyperPod
Amazon SageMaker HyperPod now includes three capabilities aimed at training jobs that stall or fail even when node hardware looks healthy. AWS describes a MultiNodeTrainingValidation deep health check that runs a short Megatron-LM training job on each instance and then across instances, failing hardware whose loss turns NaN or infinite or whose run does not finish in time. A Kubernetes resource monitoring agent on EKS clusters reports crash-looping device plugins, missing DaemonSets, tainted nodes and the pods and nodes behind a failed job. Hung job detection for Slurm watches job output and GPU activity and can cancel a stalled job so Slurm requeues it. The deep health check runs during provisioning or on demand on supported EFA-enabled NVIDIA GPU instances.
- ProductsComfyUIComfyUI 0.46.0 adds Kling, HeyGen and MiniMax partner nodes
Comfy-Org published ComfyUI v0.46.0, a release that expands the partner-node catalog and adds several core editing nodes. New integrations include LTX nodes for AlphaGen, SDR to HDR, reframing and upscaling; a MiniMax H3 Image node; Kling 4.0 Image and Long Image nodes; HeyGen Avatar 6 plus 1080p, 30-second clips; Bria RMBG 3.0 in the Remove Image Background node; and Seedance 2.5 Flash. Claude Fable 5.2 was added to the Claude node, and Wan, Krea and Luma nodes moved to Comfy Router. On the core side, the update adds a video deflicker node, a faster RAFT optical flow implementation, an image channel split node, and single-channel PNG saving. DirectML support was removed, and a CUDA 13 warning was added.
- ModelsCloudflareCloudflare releases open-weight Clairic and Clairic-flash extraction models
Cloudflare published an updated Clairic and a new Clairic-flash, both open-weight, from the group it now calls the Workers AI superintelligence team. Clairic is the 27-billion-parameter extraction model, rebuilt with a new architecture that reads text, images and PDFs — up to four images or pages per request — and still returns confidence scores, character spans and probabilities through the typed API. Clairic-flash is a 9-billion-parameter version built for lower latency. Both are served on Workers AI at @cf/cloudflare/clairic and @cf/cloudflare/clairic-flash, with full weights on Hugging Face. Clairic is priced at $0.24 per million input tokens and Clairic-flash at $0.10, though both cost $0.10 until December 1, 2026. Output tokens are not billed. Existing preview-endpoint users get the new model automatically.
- ProductsAWS Machine LearningAWS launches SageMaker JumpStart Model Leaderboard for comparing models
AWS has added a Model Leaderboard to Amazon SageMaker JumpStart, giving SageMaker Studio users a side-by-side view of open-weight foundation models on quality, speed and cost. Quality appears as separate scores from three reasoning benchmarks — BIG-Bench Hard, GPQA Diamond and MMLU-Pro — which AWS runs itself under one evaluation setup so results are comparable across models. Speed and cost figures come from a standardized test workload sent to each model's JumpStart deployment configuration, and can be narrowed by use case profile and optimization goal. At launch the leaderboard covers more than 70 models from providers including Meta, Google, Qwen, IBM and Arcee, with more to be added over time.
- ProductsAWS Machine LearningAmazon Bedrock AgentCore Identity adds private certificate authority support
AWS added private certificate authority support to AgentCore Identity, the identity capability inside Amazon Bedrock AgentCore. Teams that self-host an identity provider such as Keycloak or PingFederate on a private network can now have their agents reach it while trusting the private CA the organization already uses. The PEM-encoded CA certificate, root or intermediate and up to 16 KB, is stored in AWS Secrets Manager or Amazon S3 and referenced from a credential provider; AgentCore Identity uses it to validate the certificate presented by the provider's discovery, JWKS and token endpoints. The feature requires a private endpoint and Amazon VPC Lattice, and is available in regions where AgentCore Identity supports private identity providers. AWS published a four-step walkthrough covering certificate issuance with AWS Private CA, storage, configuration and CA rotation.
- Open sourceCrewAICrewAI 1.15.42 makes ChromaDB optional and defaults knowledge to local Qdrant
The update makes ChromaDB an optional extra and changes Knowledge storage to default to local Qdrant, separating it from ChromaDB. It also lets users pick tools for standalone agents and attach them later. The release includes a long list of bug fixes, including JSON output storage, per-Flow IDs, execution-time handling, UTF-8 file reads, and a raised soupsieve minimum for a CVE. The changes are available now to CrewAI users through the project’s release channel.
- DevelopersLangChainLangChain Anthropic package adds Claude Fable and Mythos 5.2 support
LangChain published version 1.7.11 of langchain-anthropic, its client library for Anthropic models, as a GitHub release. The update adds support for the Claude Fable and Mythos 5.2 models and sends langsmith_gateway headers from the client. It also carries fixes: a chunk is now emitted when an empty thinking block starts, refusals raise ModelRefusalError when tool calls are not forced, and the nested caller field is preserved on server calls and results. Separately, end-of-life Python 3.10 was removed from the partner packages, and the bundled model profile data was refreshed. The release notes list the merged pull requests behind each change for developers using the package.
- DevelopersNVIDIA DeveloperNVIDIA releases Daviator GPU kernels to speed up VASP plane-wave DFT
NVIDIA has introduced Daviator, a library of GPU kernels that handles the k-point steps of plane-wave DFT calculations in batches instead of sequentially, plus a mixed precision option for its batched Davidson eigensolver. Against VASP 6.6.0 CPU and OpenACC baselines, NVIDIA reports up to 22x faster time-to-solution and up to 75% lower cost for a projected OMat24 dataset workload, and more than 85% shorter execution time for a 64-atom Bi-doped GaAs system. Both capabilities are available through the NVIDIA ALCHEMI VASP microservice by setting LDAVIATOR=True and LDAVIATOR_MIXED_PRECISION=1 in the INCAR file, with existing VASP inputs running unmodified. Daviator is currently limited to single-node execution.
- ModelsVercelLiquid AI's d1.2 and d1.2-ultra decision models arrive on Vercel AI Gateway
Vercel added Liquid AI's d1.2 and d1.2-ultra decision models to AI Gateway, the company said in a changelog post. Rather than generating text, these models return structured answers: Boolean questions give yes/no probabilities, Choice questions pick a category, and Score questions rate an input against a rubric. Several questions can share one input request. Audio input and bounding-box support are included; Box questions return coordinates as fractions of image height and width. Developers call them as liquid/d1.2, the faster and lower-cost option, or liquid/d1.2-ultra, which Liquid AI positions as its more accurate decision model, using an AI Gateway API key. The audio example requires AI SDK 7.0.136 or later. Decision requests show up in AI Gateway logs and count toward budgets.
- ModelsVercelVercel adds OpenAI's GPT-Live 1.1 voice model to AI Gateway
Vercel's changelog says OpenAI's GPT-Live 1.1 is now available through AI Gateway. The model handles voice in full duplex, meaning it listens while it speaks, so a caller can pause, cut in, or add detail partway through a reply. Vercel says it transcribes speech more accurately than the earlier GPT-Live and holds a steadier accent and voice across languages. Existing GPT-Live integrations need only swap in the new model ID, since the Live endpoint and client delegation are unchanged. Developers can try it in the model playground.
- ProductsDatabricksDatabricks launches Unity Gateway Insights to find wasted AI tokens
Databricks introduced Unity Gateway Insights, a feature that uses its Genie assistant to analyze agent traces alongside a customer's business data and ontology, surfacing where tokens are wasted. The company says the tool flags recurring MCP errors, routine tasks routed to expensive frontier models, and workflows repeated across sessions that could become shared skills, ranking each finding by cost. Customers with traffic already running through Unity Gateway can open AI Gateway, then Observability, then Insights to generate a report for the past day, 7 days, or 30 days, and can add up to 10 custom questions. Databricks reports that its own trace analysis found an estimated $1.2M in annual savings from fixing seven MCP server bugs. Traces are stored in a governed Delta table in Unity Catalog.
- ProductsDatabricksDatabricks brings Genie to Microsoft Copilot, Teams and Excel
Databricks said the Genie app for Microsoft Teams is now generally available, letting users direct-message it or @mention it in a channel or group chat to get answers grounded in governed Azure Databricks data, schedule recurring analyses, and receive charts and documents in the thread. The company also described an Azure Databricks Genie plugin that connects Genie One to Microsoft Copilot Cowork and Copilot Chat, a Genie add-in for Excel that returns cited answers and live pivots on Unity Catalog tables, and a Power Platform connector for building Copilot Studio agents that call Genie Agents. Unity Catalog enforces data permissions, and Databricks says Genie Ontology, its semantic context layer, is now generally available.
- ProductsVercelVercel offers Parallel web search free on AI Gateway through January 2027
Vercel announced that Parallel's web search, in fast mode, is free on its AI Gateway through January 31, 2027, and that Parallel is now the default search provider for eve agents. The tool works with any model served through AI Gateway and does not need a separate Parallel API key. It returns links with short excerpts, and results can be filtered by website or publication date. Inside eve, the built-in web_search tool uses Parallel fast mode automatically when the model runs through AI Gateway. Developers working with the AI SDK can add gateway.tools.parallelSearch({ mode: 'fast' }) to their tools. The offer covers fast mode only, with Zero Data Retention on by default; other modes bill at standard pricing.