@TencentHunyuan on X published this on Nov 16, 2026 (source: x.com).
Tencent's Hunyuan team introduced Hy4 flash, a model it says is aimed at everyday productivity. Official API list prices are $0.003 cached, $0.070 input, and $0.281 output per million tokens. Tencent says the model matches Claude Haiku 5.5 (high) on its own daily-work human blind tests and shows an 18x task-cost reduction in those tests; those figures are the company's results, not independent evaluations. Tencent has not said the model is open-weight, and the list prices should not be described as discounted or promotional rates.
OfficialInternLM publishes Sage-S2.1-Flash on Hugging Face
Hugging Face @internlm published this on Nov 16, 2026 (source: huggingface.co).
The InternLM team published Sage-S2.1-Flash on Hugging Face. This item records that official model listing only and does not add benchmark, license, or availability claims beyond the listing itself.
Officialllama.cpp b12409 fails backend-op tests when types or backends are missing
GitHub ggml-org/llama.cpp published this on Nov 16, 2026 (source: github.com).
The llama.cpp project released b12409, making test-backend-ops fail with exit status 1 when missing or unsupported types or backends are provided on the command line (#30279).
OfficialOpenAI science blog: a step toward a lattice definition of the Standard Model
OpenAI published this on Nov 16, 2026 (source: openai.com).
In the first guest post on OpenAI's science blog, physicist Nathaniel Craig, a visiting scientist at OpenAI, describes work with GPT-6 Astra on a candidate lattice definition of the Standard Model. The post presents this as a research step, not a completed definition of the theory.
OpenAI published this on Nov 16, 2026 (source: openai.com).
OpenAI started a science blog for scientists and academic researchers to share discoveries and methods using AI. The launch post says the blog will cover research results, practical methods, and perspectives from the scientific community.
@OpenRouter on X published this on Nov 16, 2026 (source: x.com).
OpenRouter listed Qwen4-Max on its platform. This is a separate availability event from Alibaba's own Qwen4-Max production API launch and does not restate that debut, or add pricing, free-access, or open-weight claims.
OfficialGrok brings image and video creation into its unified experience
@imagine on X published this on Nov 16, 2026 (source: x.com).
Imagine's original post, corroborated by the Grok official account, describes image and video creation being brought into a unified Grok experience, so there is one place to create, build, and get work done. This is a unified entry point, not a claim that image or video generation is being introduced for the first time, and it does not add region, plan, free-access, or model-version details.
@opencode on X published this on Nov 16, 2026 (source: x.com).
OpenCode Go now includes Qwen4-Max. This covers the OpenCode Go integration only, not free access for all OpenCode users, and not an open-weight release.
OfficialAntigravity doubles the context window for Argon, Carbon, and Hydrogen
@_mohansolo on X published this on Nov 16, 2026 (source: x.com).
Varun Mohan, who leads the Antigravity team at Google DeepMind, said the context window for Argon, Carbon, and Hydrogen in Antigravity is doubling and users should see less compaction. He did not specify new token counts, plan eligibility, or regional availability.
Officialllama.cpp b12441 adds per-head flash attention masks on WebGPU
GitHub ggml-org/llama.cpp published this on Nov 16, 2026 (source: github.com).
The llama.cpp project released pre-release b12441, adding per-head flash attention mask support on WebGPU so masks with more than one head no longer produce wrong results (#32394).
OfficialAmazon Bedrock Managed Knowledge Base adds tabular search
Amazon Web Services published this on Nov 16, 2026 (source: aws.amazon.com).
Amazon Bedrock Managed Knowledge Base now supports tabular search, so agentic retrieval can answer questions over CSV, XLSX, and XLS files by looking up values and running calculations across relevant rows, then combining the results with related documents.
OfficialAWS shares best practices for Denied Topics in Bedrock Guardrails
Amazon Web Services published this on Nov 16, 2026 (source: aws.amazon.com).
AWS published eight best practices for writing Denied Topics in Amazon Bedrock Guardrails, so the guardrail matches a content category by meaning rather than keyword and blocks what was intended without overblocking.
OfficialAmazon Bedrock Managed Knowledge Base can customize answers per tenant
Amazon Web Services published this on Nov 16, 2026 (source: aws.amazon.com).
AWS showed how retrieval instructions in Amazon Bedrock Managed Knowledge Base can set the voice, format, retrieval guidance, and answer policy for each tenant on a shared knowledge base, using metadata filters, IAM tags, and per-tenant rollouts.
OfficialAWS shows prompt caching for open-weight models on Amazon Bedrock
Amazon Web Services published this on Nov 16, 2026 (source: aws.amazon.com).
AWS showed how prompt caching lets open-weight models on Amazon Bedrock reuse repeated long context to reduce input cost and latency, including implicit and explicit caching through the Chat Completions API.
OfficialOpenAI advances computer-use agents for healthcare teams
OpenAI published this on Nov 16, 2026 (source: openai.com).
OpenAI published new evaluations of computer-use agents for clinical and administrative workflows, and said healthcare teams can use the capabilities through the OpenAI API, ChatGPT for Healthcare, and ChatGPT for Clinicians. The post names work with Abridge, Ambience, Cigna Healthcare, and CVS Health. Reviewers rated 99% of 325 tasks safe, and physicians called the work very or extremely helpful in 93% of 646 assessments.
OfficialNVIDIA expands Slurm to bring quantum computing into supercomputing
NVIDIA Blog published this on Nov 16, 2026 (source: blogs.nvidia.com).
NVIDIA said quantum hardware developers, software providers, and supercomputing centers are adopting the community-built Slurm workload manager and the NVIDIA quantum platform to integrate quantum processors and develop scientific applications.
OfficialNVIDIA Vera Rubin debuts in MLPerf Training and Endpoints
NVIDIA Blog published this on Nov 16, 2026 (source: blogs.nvidia.com).
NVIDIA said the Vera Rubin platform, in its MLPerf debut with results released today, delivered leading performance across training, post-training, and agentic inference.
OfficialNVIDIA expands CUDA-X libraries for scientific discovery agents
NVIDIA Blog published this on Nov 16, 2026 (source: blogs.nvidia.com).
NVIDIA said expanded CUDA-X libraries, including cuPhoton, cuPotential, and cuSpa, give agents tools to work through large scientific datasets in areas such as astrophysics, materials science, and single-cell biology.
OfficialNVIDIA announces Rubin HPC for science, engineering, and analytics agents
NVIDIA Blog published this on Nov 16, 2026 (source: blogs.nvidia.com).
NVIDIA announced Rubin HPC, a new GPU aimed at accelerating agents for science, engineering, and analytics, saying more CUDA cores and faster data movement should shorten time to results.
OfficialOpenRouter now powers inference routing on Hugging Face
OpenRouter published this on Nov 16, 2026 (source: openrouter.ai).
OpenRouter said chat models on Hugging Face Inference Providers now run through its routing across providers, with automatic selection based on live price, speed, and reliability. OpenRouter can also send a trace of each request to a Hugging Face Storage Bucket. The company said 160 models from 38 labs are available this way.
OfficialTogether AI says docs and API fixes lifted coding-agent task completion to 97%
Together AI published this on Nov 16, 2026 (source: together.ai).
Together AI said it built agent evals to test how effectively coding agents use its platform, and that small fixes to its docs, skills, and API errors lifted task completion from 75% to 97%.
GitHub ggml-org/llama.cpp published this on Nov 16, 2026 (source: github.com).
The llama.cpp project released pre-release b12443, fixing a Vulkan Nvidia fp32 flash-attention shared-memory overflow and disabling shared-memory staging when it would use too much memory (#32244).
OfficialMoEngage built Merlin AI agents on Amazon Bedrock AgentCore
Amazon Web Services published this on Nov 16, 2026 (source: aws.amazon.com).
AWS said MoEngage built Merlin AI, a set of AI agents on Amazon Bedrock and Amazon Bedrock AgentCore that let marketing and product teams draft, analyze, and adjust campaigns from a single conversation, including turning existing APIs into MCP tools while keeping each tenant isolated.
OfficialAWS shows a path from mainframe Assembler to Java with Bedrock and AWS Transform
Amazon Web Services published this on Nov 16, 2026 (source: aws.amazon.com).
AWS published a multi-stage method that uses Amazon Bedrock, Kiro IDE, and AWS Transform to extract business rules, convert mainframe Assembler through an intermediate COBOL stage, and deploy modernized Java Spring Boot applications on AWS.
OfficialNovo Nordisk builds a manufacturing assistant on Amazon Bedrock AgentCore
Amazon Web Services published this on Nov 16, 2026 (source: aws.amazon.com).
AWS said Novo Nordisk's Manufacturing Technology team built a process-intelligence assistant on Amazon Bedrock AgentCore, using a GraphRAG architecture that pairs a knowledge graph in Amazon Neptune with hybrid search in Amazon OpenSearch Serverless to turn millions of controlled manufacturing documents into grounded, cited answers.
OfficialOpenAI expands ChatGPT for Academic Researchers to 200,000 scientists
OpenAI published this on Nov 16, 2026 (source: openai.com).
OpenAI is expanding the ChatGPT for Academic Researchers program to give up to 200,000 researchers worldwide free access to ChatGPT. Applications for the next round open Monday, November 23, with 50,000 new grants; each grant gives the recipient and up to two collaborators free ChatGPT Plus access. The company said data shared through the program is not used to train models by default.
OfficialOpenAI reports how 1,800 scientists used ChatGPT in a summer pilot
OpenAI published this on Nov 16, 2026 (source: openai.com).
OpenAI's Economic Research team studied more than 580,000 scientific messages from 1,819 scientists in the ChatGPT for Academic Researchers summer pilot and surveyed 251 participants. Surveyed participants estimated the same work would have taken 1.9 times as long without AI, and 98% of those who answered expected AI to accelerate discovery in their field. OpenAI said it published the findings in a new report.
OfficialOpenAI begins embedded assessments with Apollo Research, METR, Transluce, and Goodfire
OpenAI published this on Nov 16, 2026 (source: openai.com).
OpenAI said it has signed agreements with four independent assessors—Apollo Research, METR, Transluce, and Goodfire—for its first phase of embedded assessments. The company said the assessors will have deep access to internal models and systems, including checkpoints, source code, training and evaluation data, and monitoring systems, to assess model alignment and safeguards over an extended period.
@arena on X published this on Nov 16, 2026 (source: x.com).
Arena.ai says Qwen4-Max debuted at No. 14 on Agent Arena, reporting a 7.5% net improvement over Qwen3.8-Max across 14,700 agentic sessions. These are results reported by the evaluation platform, not a claim of universal model superiority.
Officialllama.cpp b12435 makes Vulkan radix Top-K tie selection deterministic
GitHub ggml-org/llama.cpp published this on Nov 16, 2026 (source: github.com).
The llama.cpp b12435 pre-release makes radix Top-K tie selection deterministic on Vulkan, matching the tie-selection behavior already used on CUDA and Metal. This concerns the Vulkan selection path, not a guarantee that full model outputs are deterministic.
OfficialReplit's Microsoft Fabric integration becomes generally available
@Replit on X published this on Nov 16, 2026 (source: x.com).
Replit announced general availability of its Microsoft Fabric integration. Teams can describe an app, build it with Replit Agent, and deploy it into their organization's Fabric tenant.
OfficialQwen4-Max places No. 11 overall on the Vals Index
@ValsAI on X published this on Nov 16, 2026 (source: x.com).
Vals AI reports that Qwen4-Max ranks No. 11 overall and is the highest-ranked Chinese model on its Vals Index. The evaluator says it improved over Qwen3.8-Max-0902, particularly on coding and finance tasks.
OfficialvLLM 0.34.0 adds 10 models and turns extensible KV cache on by default
GitHub published this on Nov 16, 2026 (source: github.com).
vLLM 0.34.0 adds support for 10 models, including Mistral Large 4, MiniMax-M3.1-Flash, and Nemotron-3.5-Transcribe, and turns the extensible KV cache on by default. Model Runner V1 is removed in favor of V2. The release notes cite speedups on Blackwell and MI355X and list breaking upgrade checks, including a default block size of 64, the removal of ROCm 7.2 images, and a huggingface_hub requirement of 2.0.0 or later.
OfficialAmazon Bedrock adds memory-aware Agentic Retrieve for managed knowledge bases
AWS Machine Learning Blog published this on Nov 16, 2026 (source: aws.amazon.com).
Amazon Bedrock Managed Knowledge Base and Amazon Bedrock AgentCore memory can now store conversations, extract facts, and bring long-term memory into a single AgenticRetrieveStream call. An AWS blog shows how to build a memory-aware, multi-user assistant in which each user sees only the documents they may access.
OfficialDuco rebuilds its financial data automation stack on Amazon Bedrock AgentCore
AWS Machine Learning Blog published this on Nov 16, 2026 (source: aws.amazon.com).
An AWS blog says Duco rebuilt its agentic AI stack on Amazon Bedrock and Amazon Bedrock AgentCore, using per-tenant isolation, a multi-model strategy, and an MCP skills architecture that it says cut context overhead by 90%. AWS said Duco cut reconciliation deployment time 10x.
OfficialNVIDIA introduces the DGX Spark HPC Developer Kit at SC26
NVIDIA Blog published this on Nov 16, 2026 (source: blogs.nvidia.com).
At SC26, NVIDIA introduced the DGX Spark HPC Developer Kit, a four-node cluster of DGX Spark systems linked in a ring over ConnectX-7 with no network switch. NVIDIA said a four-node kit provides 512GB of combined memory and 20TB of NVMe storage, and that with NVFP4 researchers can deploy models with up to 700 billion parameters locally across the four systems. An HPC SDK installation playbook is coming soon.
OfficialKotoba joins ElevenLabs to bring Japanese voice research to ElevenAgents
@ElevenLabs on X published this on Nov 16, 2026 (source: x.com).
ElevenLabs announced that Japanese voice AI lab Kotoba is joining the company. It plans to bring Kotoba's Japanese voice research to ElevenAgents to help businesses serve customers in Japanese.
OfficialNVIDIA, IBM, and Fermilab develop a quantum-classical codesign platform
NVIDIA Newsroom published this on Nov 16, 2026 (source: nvidianews.nvidia.com).
NVIDIA announced a quantum-classical codesign collaboration with IBM and Fermilab, connecting DGX B200 systems with IBM quantum computers for error-correction and applications research. It also introduced the TCBP-JOBS decoding algorithm in CUDA-Q.
OfficialNVIDIA announces a Vera CPU rack-scale platform for HPC
NVIDIA Newsroom published this on Nov 16, 2026 (source: nvidianews.nvidia.com).
NVIDIA announced a rack-scale HPC platform combining 192 Vera CPUs with Quantum-X800 InfiniBand for scientific simulation and agentic science. Systems from Dell, HPE, Lenovo, and Supermicro are expected in early 2027.
OfficialOpenAI releases a free child-safety detection tool for vetted organizations
@OpenAINewsroom on X published this on Nov 16, 2026 (source: x.com).
OpenAI announced a free tool for vetted organizations to detect child sexual abuse material, including AI-generated images. It developed the classifier with assistance from NCMEC and is working with the Tech Coalition and ROOST to broaden access.
OfficialPerplexity launches Frontier Research for autonomous research systems
@perplexity_ai on X published this on Nov 16, 2026 (source: x.com).
Perplexity launched Frontier Research, a research effort focused on multi-agent orchestration, reinforcement learning, continual learning, systems, and pre-training. It says the work will be supported by more than 10,000 GPUs and plans to publish papers, models, data, and benchmarks openly.
OfficialChatGPT for Healthcare adds more than 20 plugins and improves clinical search
OpenAI published this on Nov 16, 2026 (source: openai.com).
OpenAI introduced more than 20 healthcare plugins and improved trusted clinical search in ChatGPT for Healthcare and ChatGPT for Clinicians. The plugins are available in ChatGPT for Healthcare, subject to workspace administration; select plugins are also available to individual clinicians.
OfficialNVIDIA introduces Nemotron for Science with an open chemistry dataset
NVIDIA Newsroom published this on Nov 16, 2026 (source: nvidianews.nvidia.com).
NVIDIA introduced Nemotron for Science, combining Nemotron 3.5 Super VL with scientific evaluations, an open chemistry dataset, and post-training recipes. LitmusBench v0.2 contains nearly 50,000 chemistry reasoning questions for training and evaluation.
OfficialCataX AI raises RMB40 million seed round incubated and solely funded by Qiming
Qiming Venture Partners (in Chinese) published this on Nov 16, 2026 (source: qimingvc.com).
Qiming Venture Partners said CataX AI has completed a RMB40 million seed financing round, incubated and exclusively funded by Qiming. The Shanghai-based company is building an agent-native clinical development organization. It plans to license or co-develop preclinical and early clinical assets and advance them to clinical proof of concept, with human experts retaining key decisions.
OfficialLLM 0.37.3 fixes logging for Homebrew SQLite 3.54.1
Simon Willison published this on Nov 16, 2026 (source: simonwillison.net).
Simon Willison released LLM 0.37.3, a bug-fix release that repairs logging after Homebrew shipped SQLite 3.54.1, which changed how json_object handles nested json_array selects.
GitHub Changelog published this on Nov 16, 2026 (source: github.blog).
Enterprise administrators can now require selected GitHub Copilot customizations — skills, custom agents, hooks, MCP servers, and executable extensions — to come from plugins or managed settings rather than standalone user or project configuration.
Officialllama.cpp b12451 moves the web UI to Llama Chat
GitHub published this on Nov 16, 2026 (source: github.com).
llama.cpp prerelease b12451 moves the web UI to ui/apps/chat and names it Llama Chat, as a first step toward splitting the UI into a shared library and apps (#32032).