Today in 12
10 sources · 7 first-party announcements · 3 new models · 7 briefs
01 Model releases and updates
- 1Google DeepMind Introduces Gemini 4 Neon for Coding, Analysis, and ScienceLead storyFirst-partyGoogle DeepMind
- Cursor Makes Gemini 4 Neon AvailableCursor · X
- 2Hugging Face Open-Sources Rumble-2-Flash, a 0.5B Speech ModelFirst-partyHugging Face Blog
- 3RadixArk Releases DaisyRose-27B, Trained for More Efficient Post-TrainingFirst-partyHugging Face Blog
- 4Tencent Makes Hy4 Flash Open-Weight in FP8 and Base VariantsTencent Hunyuan
02 Product releases and updates
- 5Google DeepMind Rolls Out a New Project Genie ExperienceFirst-partyGoogle DeepMind
- 6Google Brings the Gemini App to Lower-Memory Android DevicesFirst-partyGoogle Blog: AI
- 7AWS Adds Workload Recovery and Infrastructure Checks to SageMaker HyperPodFirst-partyAWS Machine Learning Blog
- 8AWS Adds NVIDIA Dynamo Support to SageMaker HyperPod InferenceFirst-partyAWS Machine Learning Blog
03 Industry
04 Tips and perspectives
05 Briefs
- How Expert Streaming Brings Transformers MoE Inference Near Strata on a 32 GB MacHugging Face Blog
- Amazon SageMaker AI Adds Cross-Region Inference RoutingAWS Machine Learning Blog
- Perplexity Rolls Out Teammates and Multiagent on iOS and AndroidAravind Srinivas · X
- OpenAI Shares Background API Settings for Lower Per-Token PricesOpenAI Developers · X
- Greg Brockman Says GPT-6.2 Is Incredible at Browser UseGreg Brockman · XFollow-up
- GPT-6.2 Astra for Law Leads on Legal IntelligenceGreg Brockman · X
- Perplexity Introduces Context API for Webpage ContentAravind Srinivas · X
Model releases and updates

Hugging Face Open-Sources Rumble-2-Flash, a 0.5B Speech Model
Hugging Face open-sourced Rumble-2-Flash under the Apache 2.0 license. The 0.5B-parameter speech recognition model supports more than 40 languages and runs in Transformers and vLLM without custom code. On a 302-hour call-center benchmark, it reached 7.86% WER, close to NVIDIA Nemotron 3.5 Transcribe (30B) at 7.78%.

RadixArk Releases DaisyRose-27B, Trained for More Efficient Post-Training
RadixArk released DaisyRose-27B, its first model, built on Qwen3.8-27B and trained with reinforcement learning using Miles. Trained on 1,200 curated agent tasks over 18 days on 144 H200 GPUs, it raises Terminal-Bench 4.1 pass@1 from 5.6% to 12.0% and scores higher than Qwen3.8-27B on every benchmark listed. Checkpoints are available under Apache 2.0.

Tencent Makes Hy4 Flash Open-Weight in FP8 and Base Variants
Tencent Hunyuan has made Hy4 flash open-weight, including Hy4 flash-FP8 and Hy4 flash-Base, under the Tencent Hy License. The company lists the model as 224B with 12B active parameters, a 1M context, and native multimodality. Weights, code, and research details are linked on Hugging Face, GitHub, and Tencent’s research site.
Product releases and updates

Google DeepMind Rolls Out a New Project Genie Experience
Google DeepMind is rolling out a new Project Genie experience to eligible Google AI Ultra subscribers, featuring first-time access to Genie 4, upgrades to Genie 3, a creation agent, and a community gallery. Genie 4 supports real-time world changes from prompts and native audio; Genie 3 sessions can run for up to 3 minutes per generation, up from 60 seconds. Google AI Pro members can join the waitlist.

Google Brings the Gemini App to Lower-Memory Android Devices
Google is bringing the Gemini app to eligible Android Go devices after re-architecting it to run on hardware with less memory. Users can talk with Gemini Live in more than 45 languages, point a camera at notes or forms to learn and translate, and create custom images with Nano Banana. The app is available on Google Play and will come preloaded on upcoming partner devices.

AWS Adds Workload Recovery and Infrastructure Checks to SageMaker HyperPod
AWS introduces three SageMaker HyperPod capabilities for catching training failures: MultiNodeTrainingValidation checks whether GPUs train correctly, a Kubernetes monitoring agent reports infrastructure problems and affected jobs, and hung job detection identifies stalled Slurm jobs. Hung job detection can cancel a job so Slurm requeues it. The training check is for supported EKS clusters, while the monitoring agent covers EKS and hung job detection is available for Slurm.

AWS Adds NVIDIA Dynamo Support to SageMaker HyperPod Inference
Amazon SageMaker HyperPod Inference now supports NVIDIA Dynamo, an open source distributed inference framework for multi-node and disaggregated serving. AWS demonstrates deployment paths using DynamoGraphDeployment or InferenceEndpointConfig, and benchmarks Qwen3-235B-A22B-FP8, where disaggregated serving reached 207 output tokens per second per GPU versus 101 for aggregated serving under the stated latency and generation-speed targets. A separate Qwen3.5-397B test retained 99.1 percent of single-node throughput across two instances.
Industry
Anthropic Prices Its IPO at $83 per Share
Anthropic priced its IPO at $83 per share, with trading expected to begin on Nasdaq tomorrow. The company says the registration statement for its Class A common stock was declared effective by the Securities and Exchange Commission on November 18, 2026.

Reuters Finds Meta’s Muse Still Accessing Private Messages Across Devices
Ars Technica reports that Reuters found Meta employees had raised concerns about Muse surfacing unexpected messages and sensitive-looking financial details during testing. Meta said many suspected account mix-ups were hallucinations caused by a bug it had fixed, but Reuters said it could not establish whether the financial details were real. In tests with Signal and the AI Now Institute, Reuters found Muse could read Signal messages on Android and retrieve messages on macOS, while requests were blocked on iPhone.
Tips and perspectives
Greg Brockman Says OpenAI Models Advanced AMD Chip Design in Under Five Months
Greg Brockman says OpenAI models achieved roughly an additional year of progress on an AMD chip in less than five months, with some results included in a design recently taped out. The quoted OpenAI post reports up to 14% PPA improvement on flagship GPU IP and says gains on Jalapeño, its first custom inference chip, were verified on real silicon.

Tencent Hunyuan Announces a Reddit AMA on Building Its Everyday Model
Tencent Hunyuan says its research team and Shunyu Yao will host a Reddit AMA on Friday, Nov 20, from 8–10 AM PT. They plan to discuss how they built what the post calls an ultra-fast, lowest-cost everyday model.
Tomorrow at 08:00, a briefing like this arrives in your inbox.
The same kind of briefing as the one above: a few stories worth reading. Free. Mail starts after you confirm, and every letter has a one-click unsubscribe. Read a past issue · 2026-11-18
