The latest AI news, launches, models, companies and research — collected from official sources, research feeds and reporting, grouped into stories and ranked by the BharatHunt Trend Score.
Ingestion is behind — last completed run was 18 hours ago.
News ingestion is temporarily unavailable.
The pipeline has not completed a run recently, so what is below may be behind. Nothing here is ever filled in with placeholder news.
14 stories
ReplaySSM: Cache SSM Inputs, Not State. The story centres on vLLM. Reported by arXiv. Bharat Hunt files it under AI Models — the section covering a new or updated model, its capabilities, benchmarks or availability. Open the original report for the full detail.
Introducing Reflection Beam: an open-weight model for agentic coding on AWS. It centres on Amazon, and also names Inference and vLLM. Reported by AWS. Bharat Hunt files it under Open Source AI and AI Agents — the section covering open-weight models, permissive licences and community releases.
Native-speed vLLM transformers modeling backend. The story centres on vLLM. Reported by Hugging Face. Bharat Hunt files it under AI Models — the section covering a new or updated model, its capabilities, benchmarks or availability. Open the original report for the full detail.
Channel Selection for Mixture-of-Experts Decoding Needs No Calibration Data. It centres on Benchmarks, and also names Gemma and vLLM. Reported by arXiv. Bharat Hunt files it under AI Models — the section covering a new or updated model, its capabilities, benchmarks or availability.
Push-based KV transfer to improve TTFT and TPOT for disaggregated inference in vLLM. It centres on Inference, and also names vLLM and Benchmarks. Reported by AWS. Bharat Hunt files it under AI Models — the section covering a new or updated model, its capabilities, benchmarks or availability.
Deploying Qwen3.8-2.4T-A95B on Amazon SageMaker HyperPod with vLLM. It centres on Amazon, and also names vLLM and OpenAI. Reported by AWS. Bharat Hunt files it under Open Source AI and AI Tools — the section covering open-weight models, permissive licences and community releases.
Build real-time voice applications with vLLM-Omni on SageMaker AI – Part 1. It centres on vLLM, and also names Amazon. Reported by AWS. Bharat Hunt files it under AI Models — the section covering a new or updated model, its capabilities, benchmarks or availability.
Improve TTFT for multi-turn agentic workloads with bi-directional KV transfer in vLLM. It centres on vLLM, and also names Amazon. Reported by AWS. Bharat Hunt files it under AI Agents — the section covering autonomous and tool-using systems, agent frameworks, MCP and orchestration.
Generate images and video with vLLM-Omni on SageMaker AI – Part 2. It centres on vLLM, and also names Amazon and Flux. Reported by AWS. Bharat Hunt files it under AI Models — the section covering a new or updated model, its capabilities, benchmarks or availability.
EAServe: Encode-Aware Disaggregated Serving for Multimodal Large Language Models. It centres on Multimodal, and also names NVIDIA and vLLM. Reported by arXiv. Bharat Hunt files it under AI Models and AI Hardware — the section covering a new or updated model, its capabilities, benchmarks or availability.
Accelerate agentic AI with NVIDIA Nemotron models on Amazon EKS and SageMaker HyperPod EKS. It centres on NVIDIA, and also names Amazon and Multimodal. Reported by AWS. Bharat Hunt files it under AI Agents and AI Hardware — the section covering autonomous and tool-using systems, agent frameworks, MCP and orchestration.
Efficient Expert-Parallel Communication on PCIe-Connected Consumer GPUs. It centres on Inference, and also names vLLM. Reported by arXiv. Bharat Hunt files it under AI Hardware — the section covering chips, accelerators, data centres, on-device inference and supply. Open the original report for the full detail.
Bharat Hunt links to original reporting and does not republish it. Headlines and links belong to their publishers; summaries and the trend ranking are ours. All AI stories
Our own 0-100 ranking, not an industry metric and not anyone else’s “trending” number. It combines how recently a story was covered (32%), how many independent publications covered it (28%), how fast that coverage is arriving right now (20%), how reliable those sources are (12%) and how many people opened the story (8%).
A story we cannot date carries no score at all rather than a made-up one, so some cards show no badge. Summaries are written by Bharat Hunt from the headline and the coverage — the original article is always the source of truth, and every link goes to the publisher.