Baidu published this on Nov 11, 2026 (source: ernie.baidu.com).
Baidu's 11 November 2026 product post says ERNIE 5.5 Flash Preview is the first post-trained model built on ERNIE 5.5 Flash Base, aimed at coding, 3D modeling, knowledge work and content creation. The API is on the Qianfan platform, and Baidu says the weights will be open-sourced under Apache 2.0 soon; it describes the release as an intermediate checkpoint while post-training continues.
OpenAI News published this on Nov 11, 2026 (source: openai.com).
In a Veterans Day post dated 11 November 2026, OpenAI says more than 20,000 transitioning servicemembers and veterans have claimed a year of ChatGPT Plus at no cost, and that OpenAI Academy and partners have hosted more than a dozen trainings for nearly 1,000 people in the military community. The company lists further workshops through late November and thanks ID.me for helping verify access.
@stripe on X published this on Nov 10, 2026 (source: x.com).
In a 10 November 2026 post, Stripe said OpenRouter is now part of the company and that, together with OpenRouter, it will help businesses explore new models and optimize token routing and usage. Stripe is building economic infrastructure for AI; the post does not disclose a purchase price. The acquisition was first announced as an agreement on 19 August 2026.
IBM Newsroom published this on Nov 11, 2026 (source: newsroom.ibm.com).
IBM's 11 November 2026 newsroom post says Nighthawk r3 will be built to execute 10,000 quantum operations across 520 programmable qubits, with 2,034 physical qubits, and that in 2027 IBM plans to link two r3 processors into a 1,040-qubit system. IBM also said clients and partners have run more than 34 billion circuits on Nighthawk r2, reporting speedups of up to 19x, and that the IBM Quantum Platform now unites quantum and classical computing in one environment.
Google DeepMind published this on Nov 11, 2026 (source: deepmind.google).
DeepMind's 11 November 2026 blog says AlphaGenome 2 is a next-generation DNA sequence model that predicts how genetic variants affect specific human cell types and states, rather than broad tissue averages. Trained on more than 100 million individual cell profiles across 68 tissues, it identified nearly twice as many variants that change gene expression as the prior model on a TenK10K comparison, and is available today for non-commercial research via the AlphaGenome API, with a private preview for commercial research. DeepMind says the models are for research use only and have not been validated for clinical or diagnostic use.
OfficialTencent CodeBuddy CLI adds per-turn usage_update events in stream-json output
Tencent published this on Nov 11, 2026 (source: codebuddy.cn).
Official CodeBuddy documentation describes a usage_update event for clients that run the CLI with --output-format stream-json, so they can see token and credit use for a turn before it finishes and restore the same totals when a session is resumed. Existing result.usage and total_cost_usd fields are unchanged, and older clients can ignore the new event.
Officialllama.cpp b12283 fixes SYCL clamp on strided tensor views
Ggml-org published this on Nov 11, 2026 (source: github.com).
ggml-org published llama.cpp b12283 on 11 November 2026 (published_at 2026-11-11T06:17:37Z). The notes say the SYCL clamp operator now uses the shared unary helper so it handles non-contiguous or strided views, matching other stride-safe unary ops.
OfficialSingapore Airlines uses OpenAI models to improve travel support and employee tools
OpenAI News published this on Nov 10, 2026 (source: openai.com).
OpenAI says Singapore Airlines uses its models and API for customer support, travel discovery, and employee productivity, and that CSAT for the airline's Kris chatbot has doubled. The company post is dated 10 November 2026.
OfficialCogent and OpenAI build agentic cyber-defense workflows for enterprises
OpenAI News published this on Nov 9, 2026 (source: openai.com).
OpenAI says Cogent combines its models with enterprise context so agents can help automatically remediate vulnerabilities. The company post is dated 9 November 2026.
OfficialAuditable Human-Agent Teams for Autonomous X-ray Experimentation
Posted on arXiv on Nov 11, 2026 (source: arxiv.org).
An arXiv cs.AI paper presents a coding-agent setup, grounded in validated scientific methods, for closed-loop autonomous X-ray experiments run with human-agent teams. The authors argue that agentic systems can lower the implementation barrier, while sustained operation on physical instruments is still largely unexplored.
OfficialEarthRecord: From Pixels to Land Change with Multimodal Agents over Four Decades of Satellite Imagery
Posted on arXiv on Nov 11, 2026 (source: arxiv.org).
An arXiv cs.AI paper introduces EarthRecord, a benchmark for whether multimodal agents can turn decades of satellite imagery into reliable land-change analysis. The authors say existing evaluations often use curated images or preprocessed inputs and leave out the raw observational work Earth observation requires.
OfficialTracing Psychometric Inference in Large Language Models
Posted on arXiv on Nov 11, 2026 (source: arxiv.org).
An arXiv cs.AI paper asks whether large language models represent psychological constructs internally or only produce outputs that look psychometric. The authors use a cross-persona setup across 14 Llama 3 and Qwen 2.5 variants, from 0.5B to 14B, in base and instruct forms.
OfficialSetPrint: Black-Box LLM Fingerprinting via Setwise Correspondence of Response Changes
Posted on arXiv on Nov 11, 2026 (source: arxiv.org).
An arXiv cs.AI paper proposes SetPrint, a black-box method for telling whether a target language model was derived from a source model using only responses to queries. It perturbs each query and compares the sets of response changes from both models.
OfficialAgent Architecture Cost-Success Trade-offs: A Controlled Evaluation across Enterprise-Style SOPs
Posted on arXiv on Nov 11, 2026 (source: arxiv.org).
An arXiv cs.AI paper compares four agent architectures on long, branching, tool-heavy standard operating procedures while holding the model, tasks, and tools fixed. The study reports cost and success trade-offs for enterprise-style workflows.
OfficialTimeOmni-o1: Bidirectionally Coupling Time Series Foundation Models and LLMs for Multimodal Reasoning and Forecasting
Posted on arXiv on Nov 11, 2026 (source: arxiv.org).
An arXiv cs.CL paper presents TimeOmni-o1, which couples time-series foundation models with large language models so they can reason over series and accompanying text together. The authors target both understanding questions about a series and forecasting its future values.
OfficialBrasilidade: Measuring Brazilian Identity for Large Language Models
Posted on arXiv on Nov 11, 2026 (source: arxiv.org).
An arXiv cs.CL paper introduces Brasilidade, a benchmark meant to address gaps in how large language models are evaluated in Brazilian Portuguese. The first released domain covers standardized tasks on Brazilian cultural and linguistic identity.
OfficialCapacity Allocation at the Source: Sparse Target Optimization for LLM Knowledge Editing
Posted on arXiv on Nov 11, 2026 (source: arxiv.org).
An arXiv cs.CL paper studies knowledge editing in large language models by computing a target hidden state for a new fact, then changing parameters so the original state maps to that target. It proposes a sparse target optimization method for allocating capacity at the source.
OfficialCatch It Where It Happens: Real Time Error Correction During LLM Reasoning
Posted on arXiv on Nov 11, 2026 (source: arxiv.org).
An arXiv cs.CL paper looks at correcting language-model reasoning errors as they happen, rather than only checking the final answer. The authors argue that a single wrong step often spreads and wastes later tokens, and they evaluate real-time correction during the trace.
OfficialDeletion Sensitivity Is a Property of the Model: A Three-Condition Measurement across Seven Open Reasoning Models
Posted on arXiv on Nov 11, 2026 (source: arxiv.org).
An arXiv cs.CL paper measures how reasoning models behave when earlier reasoning traces are deleted across multi-turn tool sessions. Using 94k ShareGPT conversations, the authors say 73 percent of user interactions span multiple turns, and they compare seven open reasoning models under three deletion conditions.
OfficialAn ensemble of graph generative models for assisted history matching of the Brugge field case
Posted on arXiv on Nov 11, 2026 (source: arxiv.org).
An arXiv cs.LG paper applies an ensemble of graph generative models to assisted history matching on the Brugge field case. The authors discuss how many geological realisations can fit production data equally well when a field covers several depositional settings.
OfficialStatPrompt-LLM: A Lightweight Statistical Refinement for Foundation-Style Time-Series Forecasting
Posted on arXiv on Nov 11, 2026 (source: arxiv.org).
An arXiv cs.LG paper proposes StatPrompt-LLM, a lightweight statistical refinement for foundation-style time-series forecasting models. It aims to expose trend, scale, and temporal dependence more clearly to large pretrained architectures adapted to sequential data.
OfficialMultimodal Graph Learning: A Survey of Models, Benchmarks, and Frontiers
Posted on arXiv on Nov 11, 2026 (source: arxiv.org).
An arXiv cs.LG survey reviews models, benchmarks, and open problems in multimodal graph learning, where nodes carry text, images, or other modalities and edges encode relations. It covers graph-centric fusion methods and related application settings.
OfficialLinear Probes to Bayesian Manifolds: Geometry of Arithmetic in Large Language Models
Posted on arXiv on Nov 11, 2026 (source: arxiv.org).
An arXiv cs.LG paper examines geometric structures recovered by linear probes from language-model activations, including claims about arithmetic. It argues that probe accuracy alone conflates several distinct claims about whether a model actually computes with the recovered geometry.
OfficialBranch-and-Commit for Frozen RL Agents: When and Where to Spend Test-Time Compute
Posted on arXiv on Nov 11, 2026 (source: arxiv.org).
An arXiv cs.LG paper studies a branch-and-commit procedure for frozen reinforcement-learning agents in environments whose states can be copied and resumed. An operator samples several continuations from the current state, scores them, and continues from the chosen branch, and the paper asks when that test-time compute helps.
OfficialvLLM describes encoder disaggregation for multimodal serving, with metadata-only requests and remote embedding transfer
VLLM Blog published this on Nov 11, 2026 (source: vllm.ai).
A 11 November 2026 vLLM blog post, credited to Inferact, InferMatrix and IBM, details encoder-prefill-decode disaggregation for multimodal serving. It covers metadata-only multimodal requests, a dedicated encoder execution path, remote embedding transfer with failure isolation, reuse of encoder outputs across instances, and lower-overhead EPD routing.
OfficialComfyUI v0.45.0 adds partner nodes for Pika, Kling, GPT Image 2.6 and Gemini Omni Flash 1.2
Comfy-Org published this on Nov 10, 2026 (source: github.com).
The 10 November 2026 GitHub release for ComfyUI v0.45.0 adds partner nodes including Pika H3 Flex, Pika soundtrack, SFX and music, Tencent HY LLM, Kling 4.0 Flash, OpenAI GPT Image 2.6, and Gemini Omni Flash 1.2 on the Omni video node. It also supports RelayVSR, makes SeedVR2 VAE slice length adapt to free memory, and includes Qwen 3.x performance work plus frontend and crash fixes.
OfficialOpenRouter launches an Extractions API in alpha to pull typed fields from text, images and PDFs
OpenRouter published this on Nov 10, 2026 (source: openrouter.ai).
An 10 November 2026 OpenRouter blog post by William Kofski says the Extractions API is live in alpha for pulling named, typed fields from text, images and PDFs, returning values with spans and confidence scores. It launches with three models: Cloudflare's Clairic, ScaleDown Extract and Inception's Mercury Decide. The blog dates the post 11/10/2026; the RSS feed lists only a date-level pubDate of 10 November 2026 00:00:00 GMT, so it is not added to Top5.
OfficialDatabricks announces a second-generation always-on Genie agent for autonomous enterprise work
Databricks published this on Nov 11, 2026 (source: databricks.com).
Databricks' 11 November 2026 announcement says the new Genie is an always-on autonomous enterprise agent that can take a goal, manage long-running work, follow up on missing inputs, and reach out when a person needs to decide, using Genie Ontology for company context and Unity Catalog for governance. CEO Ali Ghodsi said he uses it to keep up with his inbox, prepare for meetings, and run work in the background. Databricks said Genie and Genie Ontology are generally available, while Genie Search, Ontology Manager and the new always-on Genie are coming soon to preview and beta programs.
OfficialHelsing unveils Clarity, a sovereign military intelligence platform that keeps humans authorizing every strike
Helsing published this on Nov 11, 2026 (source: helsing.ai).
Helsing's 11 November 2026 newsroom post says Clarity is a Europe-built agentic intelligence platform that processes sensor, satellite, drone, signals, open-source and report data in parallel to identify adversary assets and prepare engagement packages for weapons systems. Helsing said a target is nominated only when independent sources agree and an operator confirms, and that every decision is recorded with the person who made it. The company did not say whether Clarity is already in military service.
OfficialDeepL unveils Aure, a Voice AI model that carries tone, emotion and pace into real-time translation
DeepL published this on Nov 11, 2026 (source: deepl.com).
A 11 November 2026 DeepL press release says Aure, now powering DeepL Voice, continuously adapts translated speech to changes in a speaker's tone, emotion, emphasis and pace while preserving their recognizable voice. DeepL said the model analyzes incoming speech in real time from both the translated words and the source audio, and at launch supports English, German, Spanish, French, Italian, Japanese, Portuguese and Korean.
OfficialNVIDIA publishes an NVFP4 quantized build of Moonshot's Kimi K3.1 Pro on Hugging Face
NVIDIA published this on Nov 10, 2026 (source: huggingface.co).
NVIDIA's Hugging Face model card, dated 10 November 2026 and first seen in the feed at 2026-11-11T00:57:46+00:00, releases Kimi-K3.1-Pro-NVFP4, a Model Optimizer quantization of Moonshot AI's 2.8T-parameter, 104B-active multimodal MoE with a 1,048,576-token context. NVIDIA says the routed MoE expert weights are converted from MXFP4 to NVFP4, reports benchmark scores close to the MXFP4 baseline, and provides a vLLM serving recipe for NVIDIA Blackwell.
Officialllama.cpp b12292 adds WebGPU PAD_REFLECT_1D support
Ggml-org published this on Nov 11, 2026 (source: github.com).
The ggml-org llama.cpp pre-release b12292, published at 2026-11-11T20:15:19Z (15:15 ET), adds PAD_REFLECT_1D support in the WebGPU backend and removes a redundant contiguity check for that operator.
Officialllama.cpp b12291 allows dequantized KV optimization with Vulkan coopmat2
Ggml-org published this on Nov 11, 2026 (source: github.com).
The ggml-org llama.cpp pre-release b12291, published at 2026-11-11T19:52:29Z (14:52 ET), enables the KV-cache dequantization optimization with Vulkan coopmat2, guarded by a heuristic based on query tokens and the GQA ratio, and adds flash-attention test cases.
Officialllama.cpp b12290 allocates the Vulkan device-fault vendor binary buffer
Ggml-org published this on Nov 11, 2026 (source: github.com).
The ggml-org llama.cpp pre-release b12290, published at 2026-11-11T19:32:52Z (14:32 ET), allocates the Vulkan device-fault vendor binary buffer and drops a redundant vector include.
Officialllama.cpp b12287 uses MMVF for small-row BF16 matrices at small batch sizes
Ggml-org published this on Nov 11, 2026 (source: github.com).
The ggml-org llama.cpp pre-release b12287, published at 2026-11-11T18:15:36Z (13:15 ET), uses the MMVF CUDA path for BF16 matrices with few rows at small batch sizes, where the MMF tile of 32 rows per block would leave most SMs idle, for example on Qwen4exp hyper-connection projections.
OfficialGoogle DeepMind launches the Gemini Decisions API and Gemini Decision 1 for fast multimodal choices
Google published this on Nov 11, 2026 (source: blog.google).
A 11 November 2026 Google blog post, published at 18:00 UTC (13:00 ET), introduces the Gemini Decisions API, powered by the new Gemini Decision 1 model (gemini-decision-1) in Google AI Studio. The API scores up to 1,000 independent yes/no, multiple-choice or rubric questions in a single forward pass across text, images, audio, video and PDFs (up to 256K tokens), returning probabilities and confidence scores in 150-250 ms end-to-end at $0.10 per 1M input tokens with zero output tokens billed. Google says it is meant for bounded choices such as tool selection, policy checks, model routing and context grading.
OpenAI published this on Nov 11, 2026 (source: github.com).
OpenAI published Codex rust-v0.167.2 (release name 0.167.2) on GitHub on 11 November 2026 (published_at 2026-11-11T22:44:46Z). The release says Codex now sends tool updates incrementally by default on Google Cloud Vertex AI, and that the prior behavior can be restored by setting incremental_tools_legacy_gcp = false under [features] in config.toml.
OfficialAWS announces a standalone Web Search API on Amazon Bedrock AgentCore
AWS published this on Nov 11, 2026 (source: aws.amazon.com).
An AWS Machine Learning blog dated 11 November 2026 (feed published_at 2026-11-11T23:13:16Z) announces the Web Search API on Amazon Bedrock AgentCore. AWS says application code can search the web with a single signed InvokeWebSearch call, with no gateway or other resources to create first, returning ranked results from a web index that Amazon operates, each with a text extract and, when available, the page title, URL, and publication time. The post says the API complements AgentCore Gateway, is signed with standard IAM credentials, and runs on the same search infrastructure that powers Alexa+, Kiro, Quick, and other Amazon products.
OfficialElevenLabs Summit announces an Interaction Engine, an ElevenAI preview, and ElevenAgents Routines
@ElevenLabs on X published this on Nov 11, 2026 (source: x.com).
ElevenLabs' official account said on 11 November 2026 at 20:17 UTC that at ElevenLabs Summit NYC it announced agent features including an Interaction Engine that picks the right LLM for each moment of a conversation, and that in the coming months enterprises will be able to work with ElevenLabs to continually fine-tune its text-to-speech and speech-to-text models, as well as open-source LLMs, on their own data. In related official posts the same day, ElevenLabs previewed ElevenAI, a work assistant that it says can dictate into any app, take meeting notes, pass context between agents and people, and provide live assistance during calls, and said Routines in ElevenAgents can handle work before and after conversations, such as pre-screening a mortgage applicant or texting customers about a confusing bill.
OfficialGoogle AI Studio says Gemini API Managed Agents can use a self-hosted sandbox
@GoogleAIStudio on X published this on Nov 11, 2026 (source: x.com).
Google AI Studio's official account said on 11 November 2026 at 20:10 UTC that people building with Managed Agents on the Gemini API can now bring their own sandbox with self-hosted environments.
OfficialGoogle says the Gemini app is now available on Apple CarPlay
@Google on X published this on Nov 11, 2026 (source: x.com).
Google's official account said on 11 November 2026 at 19:14 UTC that the Gemini app is now available directly on Apple CarPlay, and that iOS users can have hands-free conversations with Gemini Live from the car's display, including catching up on a schedule, talking through an inbox, or brainstorming out loud.
OfficialPerplexity publishes red-team research on its SPACE sandbox for Computer agents
@perplexity_ai on X published this on Nov 11, 2026 (source: x.com).
Perplexity's official account said on 11 November 2026 at 18:21 UTC that it is publishing new research from red-teaming SPACE, its VM sandbox for Computer agents. The post says 12 models were given two tasks: escape the VM to take a credential from the host, and bypass SPACE's network policy to reach a blocked URL. Perplexity says the standard Capture the Flag setup used in red-teaming can steer agents away from the security boundary being tested, and that security-specialized models do not guarantee better results.
OfficialElon Musk says Tesla self-driving is now approved in Finland
@elonmusk on X published this on Nov 11, 2026 (source: x.com).
Elon Musk said on 11 November 2026 at 17:11 UTC that Tesla self-driving is now approved in Finland, quoting a Tesla Europe post from earlier the same day (06:18 UTC).
OfficialGoogle Labs introduces Tabletop, a shared space for custom group tools
@Google on X published this on Nov 11, 2026 (source: x.com).
Google's official account on 11 November 2026 at 16:57 UTC introduced Tabletop, a collaborative space from Google Labs where people can create custom tools to build, plan, and work with friends, family, or a club. The post says a group can start a shared hub and contribute at the same time, and gives examples of live dinner polls, trip itineraries, expense splitters, and sports brackets tailored to the group.
OfficialInfron adds Kimi K3.1 Pro Cyber and HY4 Uncensored
@InfronAI on X published this on Nov 10, 2026 (source: x.com).
Infron's official account said on 10 November 2026 at 23:48 UTC that Kimi K3.1 Pro Cyber, described for vulnerability research, and HY4 Uncensored, described for visual triage and remediation, are now on Infron.
OfficialCursor rebuilds Projects so teammates and their agents can work in one workspace
@cursor_ai on X published this on Nov 10, 2026 (source: x.com).
Cursor's official account said on 10 November 2026 at 23:20 UTC that agents can now coordinate with teammates' agents, and that Projects has been rebuilt for multiplayer so every teammate and their agents can work together in one unified workspace.
OfficialOpenAI will retire Advanced and Standard voice from ChatGPT Voice on 11 December 2026
@OpenAINewsroom on X published this on Nov 10, 2026 (source: x.com).
OpenAI's newsroom account said on 10 November 2026 at 22:07 UTC that Advanced and Standard voice will be retired from ChatGPT Voice on 11 December 2026. The post says OpenAI introduced GPT-Live in July, that voice usage has nearly doubled since then, and that virtually all people using voice in ChatGPT now use GPT-Live.
OfficialGemini 4 Carbon is now generally available in GitHub Copilot
@code on X published this on Nov 10, 2026 (source: x.com).
GitHub's official @code account said on 10 November 2026 at 21:03 UTC that Gemini 4 Carbon is now generally available in Visual Studio Code with GitHub Copilot, describing it as designed for complex coding, agentic, and reasoning tasks, with a special focus on security work.
OfficialElevenLabs is bringing Memory to ElevenAgents
@ElevenLabs on X published this on Nov 11, 2026 (source: x.com).
ElevenLabs’ official account said on 11 November 2026 at 17:40 UTC (12:40 ET) that it is bringing Memory to ElevenAgents, so agents remember what gets mentioned in passing and businesses can act on it. This is a separate official post from the same day’s already-published Summit announcements of the Interaction Engine, ElevenAI preview, and ElevenAgents Routines.
Officialllama.cpp b12307 prevents out-of-bounds reads in the Metal SOLVE_TRI kernel
Ggml-org published this on Nov 11, 2026 (source: github.com).
The ggml-org llama.cpp pre-release b12307, published at 2026-11-12T04:10:18Z (11 November 23:10 ET), prevents out-of-bounds reads in the Metal SOLVE_TRI kernel (#32103).
Officialllama.cpp b12306 adds Vulkan support for the batched tensor copy API
Ggml-org published this on Nov 11, 2026 (source: github.com).
The ggml-org llama.cpp pre-release b12306, published at 2026-11-12T03:45:58Z (11 November 22:45 ET), adds Vulkan support for the batched tensor copy API and allocates a staging buffer only for copies that need it (#31418).
Officialllama.cpp b12304 reuses graph result instances across scheduler reserves
Ggml-org published this on Nov 11, 2026 (source: github.com).
The ggml-org llama.cpp pre-release b12304, published at 2026-11-12T03:19:13Z (11 November 22:19 ET), allocates graph-result instances once in the llama_context constructor and reuses them across scheduler reserves instead of recreating them (#32089).
Officialllama.cpp b12303 adds Jinja support for chained comparisons and argument unpacking
Ggml-org published this on Nov 11, 2026 (source: github.com).
The ggml-org llama.cpp pre-release b12303, published at 2026-11-12T02:55:34Z (11 November 21:55 ET), adds Jinja support for chained comparisons and argument unpacking (#32072).
Officialllama.cpp b12302 lets TTS models provide a custom text normalizer via Jinja
Ggml-org published this on Nov 11, 2026 (source: github.com).
The ggml-org llama.cpp pre-release b12302, published at 2026-11-12T02:22:59Z (11 November 21:22 ET), allows a TTS model to provide a custom text normalizer via Jinja (#32053).
OfficialDevin’s 11 November notes add customizable sidebar shortcuts, Design Mode annotations, and grouped repo setup suggestions
Cognition published this on Nov 11, 2026 (source: docs.devin.ai).
Cognition’s Devin release notes dated 11 November 2026 say users can change Option-hold sidebar session-navigation keys in Settings, filter sessions by service user, see stacked pull requests more clearly, open a compact Design Mode prompt bar with per-property editing, spotlight or pencil annotations, an eyedropper, and CSS-variable controls, approve adjacent repository setup suggestions together, mark the open session unread, and bulk-add secrets from shell-style lines.
