Thursday, Nov 12

Ggml-org published this on Nov 12, 2026 (source: github.com).

The ggml-org llama.cpp pre-release b12317, published at 2026-11-12T06:58:35Z (12 November 01:58 ET), adds an OpenCL binary GEMM kernel for Adreno A7 Q4_0 non-MoE workloads (#30947).

Ggml-org published this on Nov 12, 2026 (source: github.com).

The ggml-org llama.cpp pre-release b12316, published at 2026-11-12T06:25:50Z (12 November 01:25 ET), returns early from the CUDA MMQ stream-k fixup when there are no MoE tiles to write, matching the main kernel (#32067).

Ggml-org published this on Nov 12, 2026 (source: github.com).

The ggml-org llama.cpp pre-release b12314, published at 2026-11-12T05:22:52Z (12 November 00:22 ET), adds decomposed flash attention for Adreno on OpenCL, including a first-call build of the FA MQ kernel and direct PV matmul that avoids a V transpose (#32046).

Ggml-org published this on Nov 12, 2026 (source: github.com).

The ggml-org llama.cpp pre-release b12311, published at 2026-11-12T05:02:04Z (12 November 00:02 ET), keeps the Discover sidebar's Downloading section visible during search, prefers repos from a base model's org in results, and throttles per-file download progress so parallel file updates stay accurate (#31467).

Ggml-org published this on Nov 11, 2026 (source: github.com).

The ggml-org llama.cpp pre-release b12308, published at 2026-11-12T04:32:11Z (11 November 23:32 ET), moves CI test models to ggml-org-dev and updates the related test spec (#32096).

Ggml-org published this on Nov 12, 2026 (source: github.com).

The ggml-org llama.cpp b12318 release, published 12 November 2026 at 11:36 UTC (06:36 ET), adds a CUDA debug buffer pool and a compute-sanitizer comment (#31919). It is a point release after b12317, which is already in today's issue.

OpenAI News published this on Nov 12, 2026 (source: openai.com).

An OpenAI customer story, timestamped 12 November 2026 at 09:30 UTC (04:30 ET), says Holiday Extras is using Codex to turn customer expertise into better apps and higher booking conversion. This is a customer case study, not a new Codex product release.

OpenAI News published this on Nov 11, 2026 (source: openai.com).

OpenAI's Children First declaration page, timestamped 12 November 2026 at 00:00 UTC (11 November 19:00 ET), says the company is committing $2 million to research on AI and children's safety and to shared tools for measuring risks and opportunities for young people.

Hugging Face Blog published this on Nov 11, 2026 (source: huggingface.co).

In a Hugging Face post published 11 November 2026 at 05:01 UTC (00:01 ET), Tether AI Research introduced BotPsy, a 0.8-billion-parameter model for function calling and agentic workflows, built on Qwen3.5-0.8B and intended for on-device use on low- to medium-difficulty tasks. Benchmark figures, including results versus models up to 1.2B and some comparisons with Qwen3.5-2B-Think, are the authors' own evaluations and do not establish that it comprehensively outperforms 2B-class models. The post describes a tool-use model, not a shipped complete personal-assistant application.

QwenLM published this on Nov 12, 2026 (source: github.com).

The QwenLM qwen-code v1.0.0-preview.1 prerelease, published 12 November 2026 at 14:25 UTC (09:25 ET), fixes Qwen subscription accounts being unable to use subscription models after sign-in, so the model catalog follows sign-in state without manual provider configuration. It also syncs the subscription model catalog and thinking-budget mapping with Qwen Code 0.27.x, adds clearer error messages, and lets Qoder models use parallel tool calls.

Recraft published this on Nov 13, 2026 (source: recraft.ai).

A Recraft blog post dated 12 November 2026 says MiniMax H3 Image, H3 Image Highspeed, H3 Highspeed, and H3 Lite are now in Recraft Studio. The two image models generate and edit images at 2K or 4K, and the two video models generate 4-to-15-second clips at 480p or 768p with sound. Recraft says it checked the settings and prices in Recraft Studio on 12 November 2026. MiniMax released H3 Image on 10 November 2026, which is already covered in earlier items. Exact original time of the Recraft post was not published, so this item is not ranked.

Langchain-ai published this on Nov 12, 2026 (source: github.com).

The langchain-ai langgraph cli==0.4.43 release, published 12 November 2026 at 16:20 UTC (11:20 ET), adds Agent Server flags with hidden deployment aliases and removes a discontinued local debugger image. It also includes a minor-and-patch dependency bump.

AWS published this on Nov 12, 2026 (source: aws.amazon.com).

An AWS Machine Learning blog post, published 12 November 2026 at 16:16 UTC (11:16 ET), describes an amazon-bedrock skill that helps migrate a customer-managed Amazon Bedrock knowledge base to a Managed Knowledge Base. The coding agent plans the move with read-only discovery, can create the new knowledge base or an infrastructure-as-code template alongside the original, and validates retrieval before cutover.

AWS published this on Nov 12, 2026 (source: aws.amazon.com).

An AWS Machine Learning blog post, published 12 November 2026 at 15:40 UTC (10:40 ET), describes exposing an Amazon Bedrock Managed Knowledge Base to AI agents as MCP tools through Amazon Bedrock AgentCore Gateway while keeping per-user access control. A request interceptor injects verified identity, Policy in AgentCore validates it independently, and a resource-based policy restricts direct access to the knowledge base.

NVIDIA Newsroom published this on Nov 12, 2026 (source: nvidianews.nvidia.com).

The NVIDIA Board of Directors announced on 12 November 2026 that Jensen Huang has resigned as president, chief executive officer, and a member of the board, effective immediately. Colette Kress, executive vice president and chief financial officer since September 2013, will serve as interim CEO and continue as CFO while the board searches for a permanent successor and considers internal and external candidates. The board appointed Harvey C. Jones, a director who has served as chief executive officer and executive chairman of Synopsys, as its first chair. The company said operations, customer commitments, and the product roadmap are expected to continue without interruption, and it will still report third-quarter fiscal 2027 results on 17 November at about 1:20 p.m. PT.

Ggml-org published this on Nov 12, 2026 (source: github.com).

The ggml-org llama.cpp b12323 release, published 12 November 2026 at 20:01 UTC (15:01 ET), refactors Gemma 4 chat tooling so tool arguments are parsed with a schema-shaped PEG grammar that enforces key structure, and it removes older Gemma 4 workarounds.

Ggml-org published this on Nov 12, 2026 (source: github.com).

The ggml-org llama.cpp b12322 release, published 12 November 2026 at 18:59 UTC (13:59 ET), changes graph metadata reserves so unused pages are not initialized and made resident when a graph uses only part of its max_nodes capacity.

Ggml-org published this on Nov 12, 2026 (source: github.com).

The ggml-org llama.cpp b12321 release, published 12 November 2026 at 18:18 UTC (13:18 ET), makes the oneDNN flash-attention cache thread-safe on the SYCL backend and memoizes runtime failures.

Ggml-org published this on Nov 12, 2026 (source: github.com).

The ggml-org llama.cpp b12320 release, published 12 November 2026 at 17:44 UTC (12:44 ET), fixes SYCL ahead-of-time compilation that could produce garbage output above 4GB, adds an --aot build flag, and updates no-xmx and no-ssl build options.

Ggml-org published this on Nov 12, 2026 (source: github.com).

The ggml-org llama.cpp b12319 release, published 12 November 2026 at 16:34 UTC (11:34 ET), changes ggml_sycl_set_main_device() so its early-return check uses the thread-local cached device id already used by the rest of the SYCL backend, instead of locking the device manager on every graph compute.

Langchain-ai published this on Nov 12, 2026 (source: github.com).

The langchain-ai langgraph==1.2.21 release, published 12 November 2026 at 18:06 UTC (13:06 ET), fixes get_subgraphs so it matches whole namespace segments and lets an undeclared subgraph namespace be read with the parent graph. It also bumps dependencies, including security fixes for filelock and langgraph-api, and documents that each interrupt() call has its own id in UnknownInterruptError.

AWS published this on Nov 12, 2026 (source: aws.amazon.com).

An AWS Machine Learning blog post, published 12 November 2026 at 17:59 UTC (09:59 PT), announces updates to the Amazon Bedrock AgentCore runtime, the managed compute layer for running agents. AWS says each session still runs in its own hardware-isolated Firecracker microVM, and that sessions now start up to 15x faster, scale to 4x more compute — up to 8 vCPUs and 32 GB — and run up to 68% more efficiently. Those figures are AWS stated results for the next-generation runtime.

OpenAI News published this on Nov 12, 2026 (source: openai.com).

An OpenAI customer story dated 12 November 2026 at 20:00 UTC (15:00 ET) describes how Amplitude product agent Wave, which Amplitude says is built primarily on OpenAI models including GPT-6.1 Sol, GPT-6 Luna, and GPT-Image-2.5 Flare, identifies product work, hands plans to coding agents including Codex, and measures impact. Amplitude estimates Wave can help teams ship a product improvement in 1-2 days, and it reported a 160% lift in homepage suggestion-card engagement versus a control after one day on its own site. Wave is in beta for customers; the figures come from Amplitude evaluations and experiment readout as presented by OpenAI.

Google AI Blog published this on Nov 12, 2026 (source: blog.google).

Google announced on 12 November 2026 at 18:00 UTC (13:00 ET) that Gemini 4 Hydrogen is now available globally in Google Search for all users, calling it the biggest model upgrade to Search yet. Google says Search can better follow multi-part and typo-heavy questions, uses improved query fan-out to run simultaneous queries and surface a broader set of links, and returns more visual, structured answers such as images, comparison tables, and custom tools. The company says the updates apply in all languages already supported by its AI models, and that people can try the capabilities in Google Search today.

Norwest via Business Wire published this on Nov 12, 2026 (source: morningstar.com).

Norwest announced the close of Norwest 18, its largest fund at $4 billion, bringing total capital under management to $18 billion. The global fund supports venture and growth equity investments across technology, healthcare and consumer businesses in the U.S., India and Israel. Norwest says its expanding investment areas include physical and voice AI, agentic cybersecurity and healthcare AI. The $4 billion is the total multi-sector fund size, not an AI-only allocation or money already invested in AI.

@v0 on X published this on Nov 12, 2026 (source: x.com).

v0 announced that Meta's Astryx design system is now selectable for projects in v0, allowing it to build with Astryx components, patterns and templates in context. This is availability inside v0, not a new release of Meta's design system or an automatic migration of existing projects.

@claudeai on X published this on Nov 12, 2026 (source: x.com).

Anthropic announced that Claude Meetings is starting a beta rollout to Team and Enterprise customers. It supports collaborating during video meetings to create documents, whiteboards and artifacts. This is a limited beta rollout, not general availability for all users.

@Replit on X published this on Nov 12, 2026 (source: x.com).

Replit announced business purchasing and Google Ads workflows for Replit Agent. Purchases require the user to confirm the total through Link, with Link and Visa handling payments; Google Ads audiences and budgets also require user approval. This does not enable unconfirmed spending or guarantee coverage of every merchant.

@demishassabis on X published this on Nov 12, 2026 (source: x.com).

Google announced Gemini 4 Hydrogen across the Gemini app and developer tools, including the Gemini API, AI Studio and Antigravity. Antigravity access is limited to paid Pro and Ultra plans. Google lists API pricing at $1.50 per million input tokens and $7.50 per million output tokens. This item covers the model and tool rollout, separate from the already published Google Search update.

@cognition on X published this on Nov 12, 2026 (source: x.com).

Cognition announced support for Gemini 4 Hydrogen in Devin Desktop and CLI. This is a new model integration in Devin, not another release of the underlying Google model. Any benchmark results in the announcement are Cognition measurements, not independent rankings.

@cursor_ai on X published this on Nov 12, 2026 (source: x.com).

Cursor announced that Gemini 4 Hydrogen is available in Cursor. This is a Cursor integration of the Google model. CursorBench results cited by Cursor are its own measurements and should not be read as accuracy across all tasks.

@alexandr_wang on X published this on Nov 12, 2026 (source: x.com).

Meta announced that Muse can make phone calls for paid subscribers in the United States. The announcement does not establish free or global availability, or that Muse can complete requests with every institution.

@demishassabis on X published this on Nov 12, 2026 (source: x.com).

Google announced that preorders are open for Gemini-powered intelligent eyewear developed with Samsung, Gentle Monster and Warby Parker. The official site limits account eligibility to adults aged 18 or older in the United States and South Korea. Preorders are not confirmation of delivery or global availability; compatibility and subscription requirements remain subject to the official site.

@ElevenLabs on X published this on Nov 11, 2026 (source: x.com).

ElevenLabs announced an expansion of its Partner Program for enterprise AI deployments. The company says its partner network includes more than 550 partners across 44 countries; those are company-reported totals, not a claim that all partners newly joined or a guarantee of deployment outcomes. The original announcement was published on November 11.

@ElevenLabs on X published this on Nov 11, 2026 (source: x.com).

ElevenLabs opened a marketplace where enterprises can discover AI applications and contact partners to evaluate deployment. Purchasing applications directly with committed ElevenLabs credits is planned for next year, not a transaction capability already available. The marketplace is distinct from the Partner Program.

@runwayml on X published this on Nov 12, 2026 (source: x.com).

Runway announced 3D model generation with Tripo H3.1. Users can generate models from text or images, upload an existing model, select angles for image or video references, and export models. This is a Runway integration, not the first release of the underlying Tripo model or a claim that every 3D asset is editable.

@nvidia on X published this on Nov 11, 2026 (source: x.com).

ElevenLabs announced that Nemotron 3.5 Super VL is available in ElevenAgents as its NVIDIA partnership expands. An additional 15MW of Vera Rubin capacity is planned, not already in production, while on-device voice for physical AI remains exploratory. Any reported cost reduction is attributed to company testing.

AWS published this on Nov 12, 2026 (source: aws.amazon.com).

AWS published a guide to per-identity Amazon Bedrock cost visibility and enforceable spend limits using only AWS-native features, with no application code for the common case. Spend can be attributed with IAM principal tags, reported with CUR 2.0 or Cost Explorer, and capped with AWS Budgets actions that apply a Deny policy when a threshold is crossed.

AWS published this on Nov 12, 2026 (source: aws.amazon.com).

AWS says comparing per-token prices hides the real cost of running a model on Amazon Bedrock. The post shows how to calculate cost per task across hidden drivers, reduce it with caching, routing, and output constraints, and assess agent ROI with human baselines, holdout experiments, and a five-dimension scorecard.

AWS published this on Nov 12, 2026 (source: aws.amazon.com).

AWS published a pattern for a self-service LLM gateway that runs LiteLLM on AWS Fargate and puts Amazon Bedrock foundation models behind an OpenAI-compatible API. Developers get access through a single API key, while platform teams can apply per-user cost controls, quota isolation, and observability.

QwenLM published this on Nov 12, 2026 (source: github.com).

QwenLM published Qwen Code v0.27.3 with managed-agent fixes for MCP session cleanup and Harness error handling, web-shell background-task chips and a more compact approval picker, localized bound-session prompts, and managed-MCP live-only selection capacity. The README now also points to the Qwen Code 1.0 Preview.

VLLM Blog published this on Nov 12, 2026 (source: vllm.ai).

vLLM and Inferact published a November 12 post on serving DeepSeek-V4.1-Flash on one NVIDIA Vera Rubin NVL72 rack. They report 1.7 million tokens/s per GPU on SemiAnalysis AgentX at a P90 of 387 tokens/s per user, which they say is 58x the throughput of a GB300 NVL72 rack at the same interactivity.

VLLM Blog published this on Nov 12, 2026 (source: vllm.ai).

vLLM and Thinking Machines say vLLM supports Inkling 2 Small from day 0, including attention residuals, a new decode kernel for relative-bias attention, all-reduce fused with surrounding ops, and DSpark speculative decoding.

Ggml-org published this on Nov 12, 2026 (source: github.com).

ggml-org released llama.cpp b12329, a prerelease. The build fuses NVFP4 weight-scale handling into the CUDA MMQ write-back path to speed prefill on NVFP4 checkpoints.

Ggml-org published this on Nov 12, 2026 (source: github.com).

ggml-org released llama.cpp b12328, a prerelease. It limits backend prefetch to inputs already computed by earlier splits, and skips the MoE cache when its device cannot run the cached matmul, avoiding a full-bank copy on every micro-batch.

Ggml-org published this on Nov 12, 2026 (source: github.com).

ggml-org released llama.cpp b12327, a prerelease. It dispatches the SpacemiT Q4_1 32x256 pack through the existing HP kernel, completing the M1 path already used for Q4_0.

AWS published this on Nov 12, 2026 (source: aws.amazon.com).

AWS describes how Natera built SWARM, an agentic system on Amazon Bedrock AgentCore, to capture lab-operations knowledge and use it at scale. The post walks through knowledge-base construction, deployment, and evaluation; Natera and AWS report about 130x faster first response and 96.2% automated coverage across support tickets.

AWS published this on Nov 12, 2026 (source: aws.amazon.com).

AWS says an Amazon Bedrock inference bill can keep rising even after prompts and application logic are tuned. The post looks at four request-level choices — cache duration, service tier, reasoning effort, and maxTokens — that change how many tokens are processed and the rate applied, and how to test each effect.

OpenAI News published this on Nov 12, 2026 (source: openai.com).

OpenAI published a November 12 customer story on Gondola, a platform that helps creatives get credit and get hired. Gondola says Codex raised its lead engineer’s merged-change output by about 50% and team engineering output by about 25% with no added engineering headcount, and that company revenue more than doubled year over year.

Official release notes published this on Nov 12, 2026 (source: docs.devin.ai).

Cognition published its November 12 Devin update with a redesigned Review experience and separate controls for automatic fixes. Veracode and Black Duck integrations are in closed beta, not generally available.

@thinkymachines on X published this on Nov 12, 2026 (source: x.com).

Thinking Machines says Inkling 2 Small weights are now downloadable on Hugging Face under Apache 2.0. This is the weights release, distinct from the existing vLLM day-0 serving support event.