Mistral AI has launched Voxtral Transcribe 2, its next-generation speech-to-text model, comprising Voxtral Mini Transcribe V2 for batch transcription and Voxtral Realtime for low-latency applications. The former supports 13 languages, speaker separation, context biasing, and word-level timestamps, achieving a word error rate of approximately 4%, priced at $0.003 per minute. The latter targets voice agents and real-time captioning, with end-to-end latency under 200 milliseconds. The Realtime model weights are open-sourced under the Apache 2.0 license, supporting long audio processing up to about three hours and enabling compliant deployment.

NVIDIA has released and open-sourced the Nemotron ColEmbed V2 family of multimodal embedding models (3B/4B/8B), based on a late-interaction architecture with ColBERT-style fine-grained interaction, designed for mixed text-image document retrieval and multimodal RAG. Its 8B version leads the ViDoRe V3 benchmark with an NDCG@10 score of 63.42, significantly improving retrieval accuracy on 'visual documents' such as charts and tables. Built on base models like Qwen3-VL or SigLIP+Llama, it is trained via contrastive learning and hard negative mining, and is now available on platforms including Hugging Face.

Baidu has launched its Qianfan Deep Research Agent (Qianfan-DeepResearch Pro), claiming top performance on the DeepResearch Bench evaluation. The benchmark covers 100 PhD-level research tasks across 22 academic disciplines, assessing research report quality using the RACE framework—evaluating comprehensiveness, insightfulness, instruction adherence, and readability. The agent employs an Agentic architecture with a 'task understanding-planning-execution' loop, integrates Baidu Search and RAG to enhance information coverage and reliability, and reduces hallucinations through dynamic reflection and two-stage report rendering, producing multi-format reports. It is now available on the Baidu Qianfan platform.

The Irish government has released a national legislative framework—the General Scheme—for implementing the EU Artificial Intelligence Act (EU AI Act), establishing a decentralized governance model: approximately 15 sector-specific regulatory bodies will assume market surveillance and enforcement responsibilities, while a statutory central body, the 'AI Office of Ireland,' will serve as the coordinating hub and single point of contact. The framework proposes creating a national register to log prohibited AI practices and high-risk AI systems, and participation in national and EU-level regulatory sandboxes to support SMEs testing in controlled environments. Certain enforcement tools and penalties require judicial confirmation, and the legislation must align with EU timelines.

The embodied AI real-robot evaluation platform RoboChallenge has published its annual report covering tens of thousands of remote physical robot tests conducted between Q4 2025 and Q1 2026. Co-launched by Yuanli Lingji and Hugging Face, the platform has deployed a cluster of 20 robots across four mainstream robot types since its launch in October 2025. The report indicates that current VLA models perform stably on basic tasks like 'stacking bowls' or 'placing objects into boxes,' but achieve very low success rates on multi-step, fine-grained manipulation tasks such as 'organizing paper cups' or 'making sandwiches.' The best-performing model achieves around 50% success on the Table30 benchmark, with actual fine manipulation success below 15%. The team plans to release a public reference dataset of failure cases—a 'mistake collection'—to aid future development.

Cursor has introduced an open specification called 'Agent Trace' to standardize attribution and traceability of AI-generated code within version control. At its core, the specification uses JSON 'trace records' to link specific code changes to their associated conversations, contributors (human/AI/mixed), and model identifiers, addressing Git's lack of fine-grained metadata. Emphasizing traceability and debuggability rather than code quality assessment, the format is platform-agnostic—storable in git notes, standalone files, or external databases—and compatible with multiple VCS systems including Git, Jujutsu, and Mercurial. The goal is to help teams diagnose why an agent may have deviated during execution.

Alibaba has open-sourced Paimon-cpp, providing native C++ engines with Apache Paimon lake table read and write capabilities, eliminating serialization and cross-language call overhead from Java bridging. The interface layer uses the Arrow C Data Interface for unified data exchange, avoiding compatibility issues across Arrow versions. On the read side, finer-grained ReadRanges combined with 'multi-producer, single-consumer' parallel prefetching improve I/O utilization. For primary key tables, asynchronous pipelining and pre-allocation strategies optimize merge operations and format conversion. Official results show end-to-end latency reductions of approximately 1–2x compared to Java-integrated versions, with support for RowTracking and DataEvolution to meet large-scale training data consistency and schema evolution requirements.

Industry analysis indicates that as GPU rack power reaches 132kW and next-gen systems may rise to 240kW, traditional air cooling can no longer handle high-density heat loads, making direct-to-chip liquid cooling a core infrastructure requirement for AI data centers. Market projections suggest the global liquid cooling market could grow from $2.8 billion in 2025 to over $21 billion by 2032, with a CAGR exceeding 30%. The report cites metrics such as single components handling ~1600W thermal loads and per-GPU slot cooling capacity reaching ~4500W, noting that GB200 NVL72-class systems in 50MW data center deployments can yield annual savings exceeding $4 million. Hyperscalers are increasingly adopting liquid cooling in new clusters.