AI news digest — Sunday, September 6, 2026

H Company Releases NeoMME: A Family of 260M and 800M Single-Tower Multimodal Encoders That Drop the Vision Tower and Causal Decoder

Research · Global · MarkTechPost

We look at NeoMME, a family of 260M and 800M bidirectional encoders from H Company. Unlike ColPali-style retrievers, it processes multilingual text tokens and raw 32×32 image patches in a single Transformer, with no pretrained vision tower and no causal decoder. We cover the masked discrete-diffusion pretraining objective, the dual dense and late-interaction retrieval heads, and the ViDoRe v3 results where the 260M model reaches 0.523 nDCG@10. We also break down the 255× index compression, the 51.3 pages per second indexing throughput on one L40S, and the text-retrieval gaps the authors acknow

Meta FAIR Introduces AI Research Preference Models (RPMs): Ranking ML Experiments Before Spending GPU Hours

Research · Global · MarkTechPost

AI research agents can propose far more experiments than they can afford to run. Meta FAIR, Oxford and UCL introduce AI Research Preference Models — frozen LLM judges that rank 15 unexecuted candidates and execute only one. On AIRS-Bench, the average normalized score rises from 0.684 to 0.729, and the baseline's 24-hour result arrives in roughly 15 hours. The post Meta FAIR Introduces AI Research Preference Models (RPMs): Ranking ML Experiments Before Spending GPU Hours appeared first on MarkTechPost .

UC Berkeley Researchers Release CUA-Lite, an Open Platform Unifying Sandboxes, Data, Evaluation and RL for Computer-Use Agents

Research · Global · MarkTechPost

Researchers from UC Berkeley have introduced CUA-Lite, an open-source platform designed to streamline the development of computer-use AI agents. Currently, the field is hindered by fragmented formats for evaluation, training, and testing. By unifying these components under a single data schema and action space, CUA-Lite significantly reduces resource overhead—dropping virtual machine sizes from 4.1 GB to a compact 0.9 GB container. This standardizes testing protocols, making it much easier for developers to build and benchmark autonomous agents.

What this means for you: AI researchers and engineers should adopt CUA-Lite to simplify their testing pipelines and improve interoperability between agent development environments.

Perplexity Details Its GPU Embedding Stack: How Ivy, Tulip and ROSE Serve pplx-embed

Research · Global · MarkTechPost

Perplexity has provided a rare technical deep dive into the infrastructure driving its search retrieval systems. By detailing its custom GPU embedding stack—Ivy, Tulip, and ROSE—the company illustrates how it optimizes the speed and cost of running large-scale ranking models. Effective AI search requires balancing model accuracy with computational efficiency, and by building a dedicated serving layer for its 'pplx-embed' system, Perplexity is setting new standards for how real-time semantic search indices are managed at scale.

What this means for you: Engineering teams building retrieval-augmented generation systems should study Perplexity’s stack to understand how specialized GPU serving layers can reduce search latency.