Skip to content
AI Daily · Asia/Shanghai · 2026-11-19 · Thursday· Issue 57· About 13 minutes to read

Google DeepMind Introduces Gemini 4 Neon for Coding, Analysis, and Science

Google DeepMind introduced Gemini 4 Neon, a model for complex software engineering, knowledge work, and science workflows, with a 1 million token context window. Google reports that Neon ranks first on the Artificial Analysis Coding Agent Index and several other benchmarks, and costs 32% less per task than Carbon at its default high thinking level. Neon is rolling out across the Gemini API, Google AI Studio, Gemini for business, Google Antigravity, and the Gemini app, with access varying by plan.
Time machineChronicle

Google DeepMindFirst-partyReported by 6 other sourcesRead the storyOriginal

Today in 12

10 sources · 7 first-party announcements · 3 new models · 7 briefs

01

Model releases and updates

Hugging Face BlogFirst-party

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%.

Hugging Face BlogFirst-party

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.

02

Product releases and updates

Google DeepMindFirst-partyReported by 2 other sources

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 Blog: AIFirst-party

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 Machine Learning BlogFirst-party

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 Machine Learning BlogFirst-party

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.

03

Industry

Ars Technica · AI

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.

04

Tips and perspectives

Greg Brockman · X

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.

(End of issue)

e/acc DailyAutomatically edited from public sources. Every story links to its original · Daily archive

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