Direct announcements from model vendors, companies and research institutions on their official sites, blogs and accounts.
AWS Professional Services uses a four-agent pattern to reduce infrastructure-as-code development per application from three to four weeks to minutes. Built with the Strands Agents SDK and running on Amazon Bedrock AgentCore, the pattern connects source and target systems through MCP tools exposed by AgentCore Gateway.
Amazon Payments used a multi-objective contextual multi-armed bandit (LinUCB) on Amazon SageMaker AI to personalize content in its product-acquisition funnel. In a seven-week online A/B test, one audience segment achieved a high-single-digit relative increase in final funnel conversion, while another showed no improvement. The team ran one LinUCB model for each stage—application start, submission, and approval—and linearly combined their UCB scores to optimize the full funnel. It open-sourced a repository that can be tried with synthetic data.
NVIDIA NeMo Agent Toolkit (NAT) can use a custom memory provider to connect to Amazon S3 Vectors as a persistent memory layer for multi-agent systems deployed on Amazon EKS.
A complete implementation for Claude Platform on AWS (CPonAWS) lets production AWS workloads, developer laptops, and external CI/CD systems share one subscription while isolating production and development traffic through workspaces. It uses three access paths: a dedicated AI Services account with cross-account SigV4, workspace-level API keys, and OIDC federation. External environments use short-lived credentials and require no persistent keys.
Amazon Bedrock AgentCore supports building 'ambient agents' that listen to event streams such as Amazon S3 event notifications, EventBridge rules, and Lambda triggers and perform tasks automatically. When needed, they pause for human approval through a single ask_human tool and resume from where they stopped.
uniopen used supervised fine-tuning on Amazon SageMaker AI and prompt-level output optimization to adapt Amazon Nova 2 Lite to its retail review policies.
An AWS blog post provides a full tutorial for deploying a multi-agent music-production pipeline on Amazon Bedrock AgentCore Runtime Instances. Three agents handle composition, delivery, and compliance, sharing a session through the same runtimeSessionId and colocating on one GPU instance.
Amazon Bedrock Knowledge Bases added the AgenticRetrieveStream API, which breaks claims questions into subqueries, retrieves information over multiple rounds, and returns cited answers.
Amazon Web Services published a solution for automating the Amazon Textract adapter lifecycle across accounts using AWS CloudFormation and Terraform templates to promote and deploy adapters across multiple accounts.
An agent-driven synthetic-monitoring solution based on Amazon Nova Act and Amazon Bedrock AgentCore replaces Selenium and Playwright DOM-selector scripts with natural-language actions, reducing each workflow from more than 10 lines of selector code to three lines.
An AWS tutorial shows how to deploy two endpoints using the same vLLM-Omni DLC image on SageMaker AI: a real-time endpoint using FLUX.2-klein-4B for image generation and an asynchronous endpoint using Wan2.1-VACE-1.3B for image-conditioned video generation. A text prompt first produces a PNG, then the image and motion prompt are sent through Amazon S3 to the video endpoint. The generated MP4 is retrieved from S3. The example includes a command-line workflow and an optional Streamlit interface.
An AWS tutorial shows how to deploy Qwen3-TTS on Amazon SageMaker AI using the vLLM-Omni DLC. A bidirectional SageMaker streaming connection sends text in and streams 24 kHz PCM audio chunks out, with a Gradio app for testing.
Prompt engineering determines how well Amazon Quick's AI features respond to natural-language requests, spanning custom agents, automated workflows, and conversational data analysis. The article presents general principles and the structured CRISPE framework, emphasizing specificity, business context, and examples. It uses enterprise scenarios such as customer-service efficiency analysis to show how to reduce iteration and improve first-pass accuracy.
AWS breaks down prompt-writing approaches for Amazon Quick components: Quick Research needs a clear objective, audience, and focus area, with sub-questions decomposed and data sources constrained; Quick Flows should specify timing, data sources, calculations, and delivery goals, using numbered steps for complex logic and conversational iteration for revisions; Quick Sight queries should state the business question, metrics, dimensions, and visualization preferences.
To address aggregate questions spanning hundreds of contracts that semantic RAG search cannot answer, this approach uses structured extraction instead. An extraction agent powered by Claude Sonnet models pulls eight key fields from each PDF and attaches confidence scores; a lightweight Claude Haiku model independently checks the results, while Amazon Textract uses computer vision to resolve disagreements over signature detection.
A canvas extension in the GitHub Copilot app is a customizable interface. Users can type the `/create-canvas` skill in an agent session and describe what they want in natural language to generate an interface such as a kanban board, launch checklist, or dashboard, without writing code.
AWS presents an architecture for scaling MoE reinforcement-learning training on Amazon EKS with EFA and DeepEP, improving throughput by 40%. It addresses resource contention between rollout generation and policy training in large-scale RL workloads such as RLHF and GRPO, high-throughput communication across hundreds of accelerators, and the cross-node all-to-all overhead introduced by Expert Parallelism.
Using the open-source SkyRL framework on Amazon SageMaker HyperPod, GRPO post-training of the Qwen3-VL-8B vision-language model raised maze-solving success from 43.75% to over 95% on a fixed 64-maze evaluation set.
NarrateAI uses five techniques on Amazon Bedrock to achieve about 99% numerical accuracy while supporting real-time streaming responses: adaptive pipeline orchestration, cross-account multi-model failover, real-time streaming evaluation, a composite evaluation framework, and data-accuracy validation. Each addresses failure modes such as hallucinated metrics, API throttling, validation delays, and subjective language. Built on Amazon Bedrock AgentCore, the solution serves more than 4,000 AWS executives.
Qwen3-TTS-12Hz-1.7B-Base from Alibaba Cloud's Qwen team is now available on Amazon SageMaker JumpStart and can be deployed to a fully managed real-time inference endpoint. It can clone a voice from just a few seconds of reference audio and its matching text, without retraining.
Amazon SageMaker HyperPod with Qumulo's Cloud Native Qumulo (CNQ) and Cloud Data Fabric (CDF) lets training compute and datasets reside in different AWS Regions without copying the data.
AWS has released a WhisperX Deep Learning Container (DLC) that builds on OpenAI Whisper with batch inference, word-level timestamps from wav2vec2 forced alignment, and speaker diarization. It can be deployed to real-time or asynchronous Amazon SageMaker AI endpoints without building a custom image.
The GitHub Copilot app rebuilt its pull-request view to render huge PRs smoothly, including one with 2,200 files, more than one million changed lines, and over 400 inline comments. It separates document height into deterministic code geometry and dynamic-block geometry: code line heights are calculated precisely in advance, while dynamic blocks such as comments are anchored by stable keys to a file, line, and side, measured lazily, and adjusted only slightly to avoid scroll jumps.
Using the GitHub Copilot app and Copilot CLI, GitHub completely rewrote the Copilot agent runtime from TypeScript into more than 800,000 lines of production Rust. AI agents wrote most of the code, merged incrementally into main across 128 pull requests rather than in a single cutover.
Why it matters: Drawing on firsthand experience, the author reviews the 800,000-line Rust rewrite and shares agent-orchestration and review practices that can be reused.
GitHub's Japan and Korea marketing lead automated event operations with GitHub Copilot. GitHub Issues serve as work units, issue forms as registration forms, labels as triggers, and Actions as task runners. A single issue automatically sets up an event, screens applicants daily, and handles wrap-up when the event ends.
The GitHub Copilot app includes diff, terminal, and browser panels, letting users review, run, and preview AI-agent code changes without leaving the app. The diff panel marks additions in green and deletions in red, and supports accepting changes, commenting, or asking Copilot to revise them. The terminal can run project commands directly and supports switching between multiple windows.
TRL v1.14's AsyncGRPOTrainer can now train and sync only LoRA adapters to vLLM. Rank-1 adapters are just a few megabytes and can be transferred through a Storage Bucket mounted across multiple Hugging Face Jobs, without NCCL.
Mistral helped a European energy operator migrate a 40,000-line Fortran 77 reservoir simulator to C++. The codebase had no test suite or centralized documentation. The team first built a numerical-equivalence testing framework, then used a Skill.md to guide agents in exporting Fortran state snapshots and validating them in C++. A custom parser generated the call tree, and Vibe CLI coordinated hundreds of agents with Mistral OCR to fill documentation gaps.
The GitHub Copilot app supports multiple agent sessions at once, each in a separate Git worktree with its own context and no interference from others. Users can switch between sessions and resume where they left off. The sessions view shows each session's title and progress. The tutorial uses the tailspin-toys repository to run three tasks at once: a funded sort, an accessibility review, and tests.
A publicly available, low-cost recipe fine-tunes LFM2.5-350M with TRL's GRPO using about 500 samples and 100 training steps, raising the pass rate on IFStruct from 22.6% to 29.7%. Training can run on a free Colab or Kaggle GPU, while evaluation uses llama.cpp on a local MacBook. The code is open on GitHub.
A Hugging Face blogger reproduced Surya Narreddi's idea of getting a language model to paint watercolors using TRL and OpenEnv. The model writes about 150 lines of JavaScript with p5.brush to create the painting.
Sentence Transformers v6.0 adds a fourth model type, MultiVectorEncoder, supporting ColBERT-style late-interaction retrieval and providing a complete training recipe.
Papers with Code's hybrid search system is powered by Hugging Face Jobs, Storage Buckets, and Inference Endpoints. It currently maintains embeddings for more than 110,000 papers from arXiv and Daily Papers.
Hugging Face built a constraint-aware GPU allocator and compared it with a FIFO scheduler across seven benchmark scenarios. With the same hardware and workloads, GPU utilization improved by up to 33 percentage points, while priority-weighted throughput rose in every scenario by up to 105%.
A Hugging Face blog post presents a streaming-data loop with Strands Robots: the same `Robot()` records a LeRobotDataset and syncs it to a Storage Bucket, then `stream_dataset()` reads directly from the Hub for training without downloading the full dataset first. To deploy a trained checkpoint back to hardware, users only need to switch the mode to `real`.
Why it matters: The Strands Robots example breaks down a robotics data loop and shows how recording, streaming training, and deployment can share one LeRobot format.
Across 417 tasks in the AppWorld Test Challenge with the same CodeAct agent, Claude Sonnet 4.6 cost $79 in total ($0.19 per task), while GPT-4.1 cost $155 ($0.37 per task). Cache-hit rate was the main factor.
Hugging Face's “Profiling in PyTorch” series continues with a third post analyzing attention using an NVIDIA A100-SXM4-80GB profiler. It shows that a naive implementation triggers an extra memory copy. Changing `masked_fill` to its in-place form, `masked_fill_`, removes one GPU kernel per forward pass.
Part four of the PRX series describes its data strategy: pretraining data combines public and internal datasets; a VLM regenerates long descriptions of images, which are then converted into a streamable training corpus. The team switched to Qwen3-VL for on-the-fly text-latent computation during training, losing only about 3–4% throughput at 7B scale (adding roughly one day to a 30-day run). All images are encoded as JPEG at quality 92.
Mistral used its open-source coding assistant Vibe to build an autonomous agent that reads Rails source files, generates or improves RSpec tests, and runs in CI/CD without human intervention.
Why it matters: Mistral has shared the full setup for an agent that writes RSpec tests, including a three-layer design with AGENTS.md, skills, and custom tools.
The Mistral AI team investigated a memory leak in vLLM. In a Mistral Medium 3.1 deployment with graph compilation and disaggregated prefill/decode, system memory grew linearly by 400 MB per minute until the process ran out of memory hours later. Heaptrack showed stable heap usage, indicating the leak was outside the heap. The team traced it to KV-cache transfers through NIXL and UCX.
