AWS publishes guide to building ultra-low-latency voice AI apps

Mercor cut batch jobs from 12 hours to 2-3 hours after moving open-model workloads to Vercel AI Gateway, which added inference capacity and failover routing when providers hit rate limits. In its first 10 days on Gateway it routed 18M requests and 63B tokens.

StateBench grades state tracking at four difficulty levels; recurrent and hybrid models like Gated DeltaNet generalize much farther than transformers when trained directly on the tasks. Frontier models including Gemini 3.8 Flash, Claude Opus 5.5, and GPT-6 Sol and Luna still break down on harder variants as inputs grow longer.

AWS walkthrough shows how Bedrock AgentCore Gateway propagates user identity to ServiceNow APIs via on-behalf-of token exchange, so downstream calls carry the initiating user rather than the application itself.

Gen Digital used Amazon Bedrock AgentCore Web Search to power Norton Revamp's AI-curated news feed for professionals managing their online reputation. The AWS customer post details the build.

NVIDIA's open multilingual RE-USE model restores speech quality across real-world degradations, removing background noise, echo and other distortions. It ships with only 9.6M parameters, and a demo plus downloadable weights are available on Hugging Face.

Hugging Face blog post from IBM Research describes Palette, a small model trained to generate presentation decks that follow a given design language. No benchmark scores, parameter counts, or release details are given in the available snippet.

Moonshot AI's open-weight Kimi K3, described as reaching frontier-level coding and agentic intelligence at release, can now run with in-country inference in India on Bedrock, alongside existing global and US-based options.