Xiaomi open-sources MiMo Code, claims edge over Claude on long coding tasks

Artificial Analysis released AgentPerf, the first benchmark designed specifically for agentic AI workloads, showing NVIDIA's Blackwell Ultra NVL72 platform delivering 20x more agents per megawatt than Hopper-based systems. The benchmark reflects the fundamentally different performance characteristics of agentic AI, which chains dozens to hundreds of LLM calls with tool execution rather than single-turn completions. Results are based on real coding agent trajectories across 12+ programming languages, providing infrastructure providers and enterprises with direct metrics for deployment decisions.

Researchers from UC Berkeley, Princeton, EPFL, and Databricks introduced PixelRAG, a retrieval system that bypasses traditional text parsing by rendering web pages as screenshots and indexing them directly for vision-language models. Tested on 30 million Wikipedia screenshot tiles, PixelRAG improved accuracy by up to 18.1% over text-based RAG systems and reduced token costs by 10x. The approach addresses fundamental information loss in conventional HTML-to-text conversion pipelines.

NanoClaw and JFrog have launched an integration that routes autonomous AI agents through vetted software registries to block malicious code downloads. The system acts as an automated immune system, intercepting compromised packages and guiding agents to approved alternatives. The partnership offers free access for open-source users and commercial licensing for enterprises, addressing a growing security gap as AI agents autonomously install packages without human oversight.

Google researchers propose 'faithful uncertainty,' a technique that allows large language models to express qualified guesses rather than either confidently hallucinating or refusing to answer. The approach reframes hallucinations as 'confident errors' and enables models to hedge responses appropriately, preserving utility while maintaining trustworthiness. This addresses a core tradeoff in LLM deployment where eliminating factual errors typically forces models to abstain from answering questions they actually know.