Wednesday, Nov 18

58 items · Updated 8:30 PM ET

Today's digest →

Top stories
1OfficialCompany

Le Groupe Figaro and OpenAI partner to bring Le Figaro journalism to ChatGPT

OpenAI published this on Nov 18, 2026 (source: openai.com).

Le Groupe Figaro and OpenAI announced a strategic partnership to make Le Figaro journalism discoverable in ChatGPT through summaries and excerpts, with links to the original stories. The partnership covers current reporting and archive content from Le Figaro and associated publications. Le Figaro retains editorial independence and plans to introduce ChatGPT Enterprise across teams.

Source: OpenAI
2OfficialProduct

OpenCode v2.3.2 caches GPT-6.2 system prompts on Responses routes

GitHub published this on Nov 17, 2026 (source: github.com).

OpenCode v2.3.2 sends the GPT-6.2 system prompt as a leading developer message on Responses routes, including OpenAI, GitHub Copilot, Zen, Azure and OpenRouter, so it is cached from the second step instead of being re-billed as uncached input. The release also stops webfetch structured output from repeating the url and format it was called with.

Source: GitHub
Latest
OfficialCodex 0.170.0 adds WebRTC voice calls in the TUI and auto-refreshed OpenTelemetry auth headers

GitHub published this on Nov 17, 2026 (source: github.com).

Codex 0.170.0 adds WebRTC negotiation for realtime voice calls in the TUI, automatically refreshed authentication headers for OpenTelemetry HTTP exporters, and experimental support for identically named MCP servers across environments. It also fixes oversized WebSocket fallback to HTTP, MCP OAuth credential races, Cloud conversation reconnection, and several Windows and sandbox issues.

Source: GitHub
Officialllama.cpp b12503 expands OpenVINO model support and optimizes NPU, GPU and memory usage

GitHub published this on Nov 18, 2026 (source: github.com).

llama.cpp b12503 expands ggml-OpenVINO model support and optimizes NPU, GPU and memory usage, including a fix for SWA mask detection.

Source: GitHub
Officialllama.cpp b12502 uses MMQ for MoE on V100 GPUs

GitHub published this on Nov 18, 2026 (source: github.com).

llama.cpp b12502 switches MoE matrix multiplications on V100 GPUs to the MMQ CUDA path.

Source: GitHub
Officialllama.cpp b12500 decodes conv2d patch indices with fastdiv

GitHub published this on Nov 18, 2026 (source: github.com).

llama.cpp b12500 uses fastdiv to decode conv2d patch indices in the implicit GEMM conv2d CUDA kernel.

Source: GitHub
Officialllama.cpp b12499 runs RMSNorm with a runtime loop count on Qualcomm Vulkan

GitHub published this on Nov 18, 2026 (source: github.com).

llama.cpp b12499 updates the Vulkan RMSNorm kernel so it can run with a runtime loop count on Qualcomm devices.

Source: GitHub
Officialllama.cpp b12498 fixes LoRA and recurrent slot scheduling on the server

GitHub published this on Nov 18, 2026 (source: github.com).

llama.cpp b12498 fixes LoRA and recurrent handling in the server slot scheduler.

Source: GitHub
Officialllama.cpp b12497 adds tiny-voca support in mtmd

GitHub published this on Nov 17, 2026 (source: github.com).

llama.cpp b12497 adds tiny-voca support in mtmd and improves G2P handling.

Source: GitHub
Officialllama.cpp b12496 tunes HIP stream_k and MMQ for AMD GCN

GitHub published this on Nov 17, 2026 (source: github.com).

llama.cpp b12496 tunes the HIP stream_k algorithm and MMQ configs so stream_k can be used on AMD GCN GPUs.

Source: GitHub
Officialllama.cpp b12493 uses the implicit GEMM conv2d path for F32 kernels

GitHub published this on Nov 17, 2026 (source: github.com).

llama.cpp b12493 routes F32 conv2d kernels through the implicit GEMM path.

Source: GitHub
Officialllama.cpp b12492 adds an A7 Q5_K non-MoE OpenCL binary kernel

GitHub published this on Nov 17, 2026 (source: github.com).

llama.cpp b12492 adds an OpenCL binary kernel for Q5_K non-MoE workloads on A7.

Source: GitHub
Officialllama.cpp b12491 adds Qwen3-TTS ICL and reference mode in mtmd

GitHub published this on Nov 17, 2026 (source: github.com).

llama.cpp b12491 adds support for Qwen3-TTS in-context learning and reference mode in mtmd.

Source: GitHub
Officialllama.cpp b12490 fixes Qwen4exp graph reallocation from the n-gram table scale

GitHub published this on Nov 17, 2026 (source: github.com).

llama.cpp b12490 fixes Qwen4exp graph reallocation caused by the n-gram table scale.

Source: GitHub
Officialllama.cpp b12489 returns HTTP 400 when grammar and ignore_eos are both on

GitHub published this on Nov 17, 2026 (source: github.com).

llama.cpp b12489 returns HTTP 400 and avoids a sampling crash when grammar and ignore_eos are enabled together.

Source: GitHub
Officialllama.cpp b12488 disables quantized matvec fusion on Pascal GPUs

GitHub published this on Nov 17, 2026 (source: github.com).

llama.cpp b12488 disables quantized matvec fusion on all Pascal GPUs.

Source: GitHub
OfficialTencent publishes Hy4 flash base model weights on Hugging Face

Hugging Face published this on Nov 17, 2026 (source: huggingface.co).

Tencent published the Hy4 flash base model on Hugging Face, listing it as an open-weights Hy4 flash checkpoint for text generation. This is the base-model weights release, distinct from Tencent's earlier Hy4 flash API introduction and from Hy4 flash becoming available on OpenRouter.

Source: Hugging Face
OfficialAnthropic says Claude Fable 5.2 solved two open problems in black-hole physics

Anthropic published this on Nov 17, 2026 (source: anthropic.com).

Anthropic said Claude Fable 5.2 solved two open problems in theoretical physics related to black holes and quantum gravity. In the first, it completed a calculation set out more than 20 years ago, showing that a sudden transition associated with black-hole formation persists even where spacetime no longer has a definite shape. In the second, Claude Fable 5.2 found the exact limit of the simplest version of a technique used to test candidate theories of quantum gravity. Anthropic said the work does not describe our universe directly and has no practical applications; preprints are available from the announcement.

Source: Anthropic
OfficialGitHub Copilot adds connectors for Supabase, Box, Calendly and five other services

GitHub published this on Nov 18, 2026 (source: github.blog).

GitHub said Copilot users can now connect eight more services through GitHub-verified MCP servers, or connectors, in public preview: Box, Buildkite, Calendly, ClickUp, Mixpanel, Supabase, Webflow and Wix. The connectors are in public preview for all Copilot plans; Copilot Business and Enterprise admins need to enable the Editor preview features policy, and may also need to add the new connectors to an MCP allowlist. Some services also need to be enabled by an admin on their side.

Source: GitHub
OfficialGemini 4 Neon becomes generally available in GitHub Copilot

GitHub published this on Nov 18, 2026 (source: github.blog).

GitHub said Gemini 4 Neon is now generally available in GitHub Copilot for Copilot Pro+, Max, Business, and Enterprise users, with a gradual rollout. Google says the model is designed for agentic coding, multistep tool use, and knowledge work, and reports quality close to Gemini 4 Carbon at a lower token cost per task. The model is billed at provider list pricing under usage-based billing. Copilot Enterprise and Business administrators can manage access through the model policy in Copilot settings.

Source: GitHub
OfficialRecraft adds ByteDance's Seedance 2.5 Flash for faster video generation

Recraft published this on Nov 17, 2026 (source: recraft.ai).

Recraft said ByteDance's Seedance 2.5 Flash is now available in Recraft Studio as a faster, lighter tier of Seedance 2.5, with the same prompt-to-video workflow, optional generated sound, and clips of up to 30 seconds at 480p or 720p. ByteDance says Flash generates about twice as fast as Seedance 2.5; in Studio it starts at 15 credits per second of video, versus 18 to 192 for Seedance 2.5. Flash does not offer the 1080p or 4K options available on Seedance 2.5. This is Recraft's Studio availability, not a new ByteDance model announcement.

Source: Recraft
OfficialvLLM describes FlyDSL MonoKernels aimed at lower-latency agent decode

VLLM published this on Nov 18, 2026 (source: vllm.ai).

The vLLM project published a technical post on FlyDSL MonoKernels for high-interactivity decode in vLLM, dated November 18, 2026. The post argues that interactive agents are limited by decode latency at low concurrency, and describes fusing GPU kernels to cut dispatch and startup costs in the decode path. It is an official engineering write-up of the technique in vLLM, not a numbered version release.

Source: vLLM
Officialllama.cpp b12528 fixes sleep state for the v1/realtime server

Ggml-org published this on Nov 18, 2026 (source: github.com).

llama.cpp b12528, published November 18, 2026, fixes sleep state handling for the v1/realtime server path. The pre-release tag points to pull request #32567.

Source: ggml-org
Officialllama.cpp b12527 rejects Limite MUDD configs that would overflow tap indices on load

Ggml-org published this on Nov 18, 2026 (source: github.com).

llama.cpp b12527, published November 18, 2026, rejects Limite MUDD configurations that index past mudd_tap_idx when a GGUF is loaded. The change checks at hparams load time so a crafted file with n_layer times n_mudd_taps above 512 cannot read past a fixed 512-entry array during tensor load or graph build.

Source: ggml-org
Officialllama.cpp b12521 lets the WebGPU flash-attention mask broadcast over the Q batch

Ggml-org published this on Nov 18, 2026 (source: github.com).

llama.cpp b12521, published November 18, 2026, updates the WebGPU flash-attention path so the attention mask can be broadcast over the Q batch, and uses Q_ne3 for the get_alloc_size block-batch count. The pre-release tag points to pull request #32535.

Source: ggml-org
OfficialAmazon Bedrock AgentCore Identity adds support for private certificate authorities

AWS published this on Nov 18, 2026 (source: aws.amazon.com).

AWS said AgentCore Identity, a capability of Amazon Bedrock AgentCore, now supports private CA certificates so agents can connect to privately hosted identity providers such as Keycloak or PingFederate that present TLS certificates from a private certificate authority. Users store a PEM-encoded root or intermediate CA certificate in AWS Secrets Manager or Amazon S3 and reference it from a credential provider; the certificate contains only the public certificate, not its private key. AgentCore Identity uses it to validate the IdP's discovery, JWKS, and token endpoints.

Source: AWS
OfficialGemini expands to lower-memory Android devices

Google published this on Nov 18, 2026 (source: blog.google).

Google said the Gemini app is now available on more lower-memory Android devices, including eligible Android Go phones. The reworked app brings its core experience to devices with more limited memory.

Source: Google
OfficialNVIDIA says RTX Spark compact desktops from ASUS, Dell and MSI are available now

NVIDIA published this on Nov 18, 2026 (source: blogs.nvidia.com).

NVIDIA said new RTX Spark compact desktops from ASUS, Dell and MSI are available now, with more systems from Acer, GIGABYTE, HP and Microsoft on the way. The company said the desktops offer RTX acceleration and up to 128GB of unified memory for local AI, development and creative workflows, including a 20-core NVIDIA Grace CPU and a Blackwell RTX GPU. NVIDIA described them as dedicated Windows PCs for local models and always-on agents, not a replacement for the earlier RTX Spark laptop preorders.

Source: NVIDIA
OfficialMuse adds virtual clothing try-on while shopping

Meta published this on Nov 18, 2026 (source: x.com).

Meta said Muse can now show shoppers how clothes could look on them before they buy. The assistant offers the virtual try-on experience during clothing-shopping conversations.

Source: Meta
OfficialGenie 4 becomes available in Project Genie for eligible Ultra subscribers

Google DeepMind published this on Nov 18, 2026 (source: blog.google).

Google made Genie 4 available in Project Genie, adding native audio and real-time world editing. The updated Project Genie experience also includes a creation agent and a gallery. Eligible Ultra subscribers can try it, while Pro users can join a waitlist.

OfficialSalesforce and Microsoft announce Copilotforce collaboration

Salesforce published this on Nov 17, 2026 (source: sforce.co).

Salesforce and Microsoft announced Copilotforce, a collaboration intended to bring Salesforce data and actions into Microsoft Copilot. Sales Cloud in Teams is in a pilot, and the companies also described bidirectional zero-copy connectivity between Data 360 and Microsoft Fabric.

Source: Salesforce
OfficialSuno opens MCP access beyond its early-access signup

Suno published this on Nov 18, 2026 (source: x.com).

Suno said its MCP integration is now available to everyone, letting users create songs and visuals from tools such as ChatGPT and Claude. This expands availability beyond the previously announced early-access signup.

Source: Suno
OfficialGoogle opens a waitlist for Stitch Loop, a post-launch product agent

Google published this on Nov 18, 2026 (source: x.com).

Google introduced Stitch Loop as an experimental product agent that keeps working after launch. Using Gemini 4 Carbon and CodeMender, it monitors live metrics and works on improvements such as security, latency and accessibility. Access starts through a waitlist.

Source: Google
OfficialOpenAI reports chip-design gains with AMD and work with Synopsys

OpenAI published this on Nov 18, 2026 (source: openai.com).

OpenAI said its models identified up to a 14% power, performance and area improvement on AMD flagship GPU IP in less than five months, with some results included in a recently taped-out design. OpenAI also described work with Synopsys to train GPT-Synopsys.

Source: OpenAI
OfficialVercel opens applications for its Vercel X AI accelerator

Vercel published this on Nov 18, 2026 (source: vercel.com).

Vercel introduced Vercel X, a six-week AI accelerator in San Francisco scheduled between January and March 2027. Applications close on November 29.

Source: Vercel
OfficialMistral Vibe launches a desktop app for local and remote agents

@MistralAI on X published this on Nov 17, 2026 (source: x.com).

Mistral launched Vibe Desktop for macOS and Windows, bringing multiple agent tasks, diffs and pull requests into one window. Agents can work with local repositories or in a cloud sandbox; Linux support is listed as coming soon.

OfficialTencent reports Q3 revenue growth and more than 10 million WorkBuddy daily users

Tencent published this on Nov 18, 2026 (source: buff.ly).

Tencent reported third-quarter revenue of RMB207.3 billion, up 8% year over year. The company said WorkBuddy had surpassed 10 million daily active users and cloud revenue grew 35%.

Source: Tencent
OfficialQwen4-Max becomes available on TokenHarbor

Alibaba Cloud published this on Nov 18, 2026 (source: x.com).

Alibaba announced that Qwen4-Max is available on TokenHarbor. The update adds a distribution channel for the already released model.

OfficialManus becomes a Platinum Sponsor of the Godot Foundation

Manus published this on Nov 18, 2026 (source: manus.im).

Manus said it has become a Platinum Sponsor of the Godot Foundation, which supports the community-built open-source game engine. The sponsorship does not announce an AI integration into Godot.

Source: Manus
OfficialInfron adds 61 audio models across speech and realtime voice

Infron published this on Nov 18, 2026 (source: x.com).

Infron added 61 audio models to its platform: 38 text-to-speech models, 17 speech-to-text models and six realtime voice models. The catalog spans multiple providers and voice use cases.

Source: Infron
OfficialInfron adds a self-hosted DeepSeek V4 Flash 0731 endpoint

Infron published this on Nov 18, 2026 (source: x.com).

Infron said it has deployed DeepSeek V4 Flash 0731 itself and is offering access free of charge. This is a hosting and access update, not a new release of the DeepSeek model.

Source: Infron
OfficialGPT-6.2 Astra and Sol become available in GitHub Copilot

GitHub published this on Nov 18, 2026 (source: github.blog).

GitHub added GPT-6.2 Astra and GPT-6.2 Sol to Copilot for Pro+, Max, Business and Enterprise customers. Availability is rolling out gradually, and organizational access depends on administrator policies.

Source: GitHub
OfficialOpenAI updates API analytics for usage, limits and errors

OpenAI published this on Nov 17, 2026 (source: x.com).

OpenAI updated the API usage dashboard with improved usage and spend tracking, rate-limit information and visibility into limited requests. Developers can also investigate errors by model, project and API key.

Source: OpenAI
OfficialOpenAI says GPT-6 Cyber helped Calif identify a major Java flaw

OpenAI published this on Nov 17, 2026 (source: x.com).

OpenAI cofounder Greg Brockman said GPT-6 Cyber helped Calif find a major Java security flaw as part of the Patch the Planet initiative. His statement linked to a November 17 New York Times report.

Source: OpenAI
OfficialGoogle AI Studio adds custom domains for apps built in Studio

Google published this on Nov 17, 2026 (source: x.com).

Google introduced custom-domain support in AI Studio Build, allowing builders to publish branded apps with a domain managed within the same workflow. Google also said domains can be purchased directly in AI Studio.

Source: Google
OfficialNVIDIA expands supply commitments for AI infrastructure

NVIDIA published this on Nov 17, 2026 (source: nvidianews.nvidia.com).

NVIDIA announced expanded manufacturing commitments with Amkor, Foxconn, Kinsus, Samsung Electro-Mechanics, SPIL and TSMC. New and expanded agreements cover substrates, advanced packaging and production capacity for current and next-generation AI infrastructure platforms.

Source: NVIDIA
OfficialMicrosoft introduces MAI Realtime 1 in limited-access preview

Microsoft published this on Nov 17, 2026 (source: msft.it).

Microsoft introduced MAI Realtime 1, a speech-to-speech model available to try in MAI Labs and through a limited-access preview in Microsoft Foundry. Audio starts at $0.01 per minute; reasoning is billed separately at $1 per million input tokens and $5 per million output tokens.

Source: Microsoft
OfficialNVIDIA reports $115.8 billion in fiscal Q3 revenue

NVIDIA published this on Nov 17, 2026 (source: x.com).

NVIDIA reported fiscal 2027 third-quarter revenue of $115.8 billion, up 103% year over year. Data Center revenue reached $107.6 billion, up 110%, according to the company.

Source: NVIDIA
OfficialMuse adds voice access on Apple Watch and Wear OS

Meta published this on Nov 17, 2026 (source: x.com).

Meta said voice access to Muse is now available on Apple Watch and Android Wear OS devices. The update extends access to the assistant through smartwatches.

Source: Meta
OfficialChatGPT adds live dictation that transcribes speech as you talk

@ChatGPT on X published this on Nov 18, 2026 (source: x.com).

ChatGPT introduced live dictation, with speech appearing as text while a user talks. OpenAI describes the updated input experience as faster and more reliable.

OfficialPerplexity brings Teammates and Multiagent to its mobile apps

Perplexity published this on Nov 18, 2026 (source: x.com).

Perplexity brought Teammates and Multiagent to its iOS and Android apps for Max subscribers, including voice access to Teammates. This extends the previously announced desktop capabilities to mobile.

Source: Perplexity
OfficialReplit expands Microsoft collaboration with Local Work and Teams availability

Replit published this on Nov 18, 2026 (source: replit.com).

Replit announced general availability for Local Work on Windows and its Microsoft Teams integration as part of an expanded Microsoft collaboration. Publishing to Azure is in early access, while full bring-your-own-cloud support remains planned.

Source: Replit
OfficialInfron adds stablecoin top-ups through Coinbase Business

Infron published this on Nov 17, 2026 (source: x.com).

Infron announced a Coinbase Business integration that lets customers buy Infron credits with stablecoins and use them across models from multiple providers through one API.

Source: Infron
OfficialKling Image 4.0 expands to partner platforms

Kling published this on Nov 18, 2026 (source: x.com).

Kling announced Image 4.0 availability on Higgsfield, Eachlabs, Runware and fal. The partner rollout broadens access to the model, whose native 8K generation was already announced.

Source: Kling
OfficialQwen4-Max becomes available on Runware

Qwen published this on Nov 18, 2026 (source: x.com).

Alibaba said Qwen4-Max is available on Runware, including a one-million-token context window and image and video input. This adds a deployment channel for the existing model.

Part of Qwen
Source: Qwen
OfficialAnthropic prices its IPO at $83 a share, with Nasdaq trading expected on November 19

Anthropic published this on Nov 18, 2026 (source: anthropic.com).

Anthropic priced its initial public offering at $83 per share, above its previously proposed $70–$78 range. The shares are expected to begin trading on Nasdaq on November 19 under the symbol ANT, with the offering expected to close on November 20, subject to customary closing conditions.

Source: Anthropic
OfficialTencent releases Hy4 flash and FP8 model weights

Tencent published this on Nov 18, 2026 (source: x.com).

Tencent announced availability of Hy4 flash model weights, including an FP8 version. Tencent describes the model as having 224 billion parameters with 12 billion active, a one-million-token context window and native multimodality. The weights use the Tencent Hy license.

Part of Tencent
Source: Tencent