13 articles
Google DeepMind has introduced Decoupled DiLoCo, a distributed architecture for training AI models. It reduces the required network bandwidth from 198 Gbps to 0.84 Gbps across 8 datacenters and achieves 88% goodput compared to 27% with conventional methods.
vLLM published full integration of V4-Pro and V4-Flash models on the same day as DeepSeek's release, with an 8.7× smaller KV cache compared to V3.2 at one million token context. The combination of sparse attention and aggressive compression enables serving on standard GPU hardware.
Allen Institute for AI has launched OlmoEarth Studio with three model sizes (Nano, Tiny, Base) for satellite embeddings. The models achieve an F1 score of 0.84 for landscape segmentation with only 60 labeled pixels and support change detection and PCA visualization.
Apple presented five research papers at ICLR 2026 in Rio de Janeiro, with the most notable being ParaRNN — a method enabling parallel training of nonlinear recurrent neural networks with a 665× speedup over sequential approaches, scaling RNNs to billions of parameters to compete with transformers.
Linux Foundation AI & Data published a practical guide for implementing the RGAF (Responsible Generative AI Framework) across nine dimensions of responsible AI, with a catalog of 35 concrete open-source tools and alignment with NIST AI RMF, EU AI Act, ISO/IEC 42001, and OECD principles.
WorldDB is a new memory engine for AI agents that combines a vector database with a graph of nested worlds and an ontology. It reconciles knowledge at write time and prevents contradictions in agent memory, achieving 96.40% accuracy on the LongMemEval-s benchmark.
BAR (Branch-Adapt-Route) is a new modular approach to post-training from the Allen Institute for AI that enables independent training of domain experts — math, code, tool use, safety — and their combination into a unified mixture-of-experts model. Results on OLMo 2 7B: 49.1 average score, +7.8 points for math and +4.7 for code over the baseline retraining.
AMD FLy is a new training-free speculative decoding method that achieves 4.80× to 5.21× speedup on Llama-3.3-405B and 2.74× on Llama-3.1-70B through semantic acceptance of draft tokens, with accuracy above 99%, requiring no additional model training.
Gemma 4 is Google's new generation of open models in four variants: E2B for mobile devices, E4B for edge devices, 26B MoE with 3.8 billion active parameters, and 31B dense. The 31B holds third place on the Arena open-model leaderboard and reportedly outperforms models 20 times its size, while the 26B MoE ranks sixth. All models are multimodal (text, image, video, audio), support 140 languages, offer up to 256K token context, and are released under the Apache 2.0 license.
HuggingFace has published a detailed guide for fine-tuning multimodal embedding and reranker models through the Sentence Transformers library. The focus is on unifying text and images in a shared embedding space, enabling semantic search across heterogeneous data. The primary application is in RAG systems working with a mix of documents, tables, images, and scans.
HuggingFace has published a 15,000-word Transformers-to-MLX skill that uses Claude Code for porting Transformers models to the MLX-LM platform for Apple Silicon. The skill includes a test harness that independently verifies results, eliminating the problem of LLM hallucinations, and addresses the growing challenge of open-source projects where AI agents increase pull request volume by 10 times.
LangChain has built an automated documentation testing system using Deep Agents that prevents stale code examples in documentation. The system uses multiple skills that migrate inline code snippets into standalone test files, run them through GitHub Actions, and regenerate verified snippets, ensuring the documentation always reflects the actual state of the API.
Overworld has released Waypoint-1.5, a real-time video world model that generates interactive virtual environments at 720p/60fps on consumer GPUs such as the RTX 3090 and Apple Silicon Macs. The model was trained on ~100x more data than the previous version and is available open source on HuggingFace.