Hugging Face's blog argues AI's next real constraint is GPU utilisation, rather than intelligence. GPUs are billed by calendar hours but produce only in compute hours, mirroring the cost structure of grounded aircraft.
NVIDIA introduced Cosmos-H-Dreams, a real-time action-conditioned generative simulator for surgical robots. It runs through FlashDreams on one RTX PRO 6000 GPU and supports closed-loop interaction by humans and policies.
IBM Research shared its experience building model routing into agent systems on the Hugging Face blog, noting that most routers treat model selection as classification when it is actually a systems-optimisation problem.
Google DeepMind launched the ATL Saathi pilot in India, a Gemini-powered web app providing Atal Tinkering Labs educators with a 24/7 lesson preparation and training assistant, initially covering 100 schools.
The third Hugging Face PyTorch profiling instalment traces attention implementation with a profiler. Comparing naive attention with in-place masking shows replacing masked_fill with masked_fill_ removes one GPU Memcpy kernel per forward pass.
Mistral Studio now provides central management of Prompts and Skills, with version histories, clear ownership and traceability. Immutable versions, rollback, tags and audit logs make AI behaviour governable and discoverable, while Observability traces production outputs to their versions. Skills can be invoked directly from Studio as MCP servers, ensuring production executes the same governed asset.
Hugging Face announced that vLLM’s transformers modelling backend now matches or exceeds the throughput of vLLM’s handwritten native implementations across several LLM architectures. It uses torch.fx for static graph analysis and ast to rewrite source code, dynamically applying inference-related layer fusion at runtime to match custom-code performance.
Hugging Face and Amazon SageMaker AI introduced deep-link integration. Clicking Customize on SageMaker AI or Deploy on SageMaker AI on supported model pages opens SageMaker Studio with the model preloaded and environment configured automatically.
Google Research published the first large-scale real-world study in Nature Cities, redirecting under 2% of trips to alternative routes through Google Maps changes across ten US cities for six months. Median target-road speed rose around 2% and median fuel consumption rates fell 0.5–1.0%, potentially saving thousands of tonnes of CO2e per city annually.
At Build 2026, Microsoft announced Foundry Managed Compute and a Hugging Face model collection, with weekly updates to open-weight models and one-click deployment to Foundry's managed GPU platform.
As inference costs fell from around $30 per million tokens for GPT-4-class models in early 2023 to under $1 today, and below $0.10 at some providers, UC Berkeley's Aditya G. Parameswaran and colleagues proposed three directions for data systems in the agent era: For Agents, Of Agents and By Agents.
Hugging Face and SkyPilot released an integration allowing a Hugging Face Bucket or any model, dataset or Space repository to be mounted into SkyPilot tasks with an hf:// URL and an existing HF_TOKEN, running across more than 20 clouds, Kubernetes, Slurm and local environments.
Hugging Face introduced a new kernel repository type for Kernels, along with trusted publishers and code signing. By default, only trusted publishers' kernels load; other sources require explicit trust_remote_code=True.
NVIDIA's ENPIRE lets coding agents drive autonomous experiments and iteration with real robot arms. Dual YAM arms and an RTX 5090 workstation achieved 99% success on PushT, pin organisation and cutting cable ties. GPT-5.5/Codex and Opus 4.7/Claude Code traded wins, while Kimi-2.6 lagged. Eight agents found higher-scoring solutions faster.
Google Research published an architecture attaching multi-token prediction (MTP) heads to frozen Gemini Nano v3 models, accelerating on-device inference without changing backbone weights. It has shipped with the Pixel 9 and 10 series.
Google Research proposed linear elastic caching, modelling page eviction as a ski-rental problem and using shallow decision trees to predict page TTLs, dynamically resizing caches to minimise total ownership cost. After months in Spanner production, memory fell 15.5%, cache misses rose only 5.5%, TCO dropped around 5% and actual I/O cost impact was just 0.5%. It was also validated on several public cache traces.
Mistral AI added new Connectors capabilities. Admin controls for workspace- or organisation-level access and individual tool toggles are generally available, as are API keys scoped to connectors.
Mistral released OCR 4, returning bounding boxes, block-type classifications and per-page and per-word confidence alongside text extraction. It supports 170 languages and single-container self-hosting. Independent annotators preferred OCR 4 on average 72% of the time in blind evaluations of over 600 documents. It scored 85.20 on OlmOCRBench and 93.07 on OmniDocBench, though Mistral warns both benchmarks have known scoring limitations.
With Google's support, UC San Diego plans a data centre using motherboards from 2,000 retired Pixel phones to provide low-cost, low-carbon cloud computing for hundreds of students and faculty. SPEC benchmarks show 25–50 phones roughly equal one modern server. Phones form Kubernetes-managed clusters of 25–50 devices, with launch expected in autumn 2026.
Google DeepMind released experimental open-source model DiffusionGemma, generating text blocks in parallel through text diffusion and achieving up to 4 times faster inference on dedicated GPUs.
Mistral released an AI stack for industrial engineering at AI Now Summit 2026, partnering with Airbus, BMW and ASML to optimise design, simulation and production while retaining control over proprietary data and IP.
Google combined a new single-message encrypted aggregation protocol with TEEs, letting devices submit without multiple online rounds. Google receives only anonymised group insights; raw data is neither exposed nor reconstructed even within hardware protection. TEE attestation proves execution follows public code.
After incorporating Emmi AI, Mistral launched physical AI capabilities for AI-native industrial engineering with partners including ASML, Airbus, Safran and Siemens Energy. The model predicts physical fields directly from geometry and boundary conditions in seconds through one forward pass on a single GPU, versus hours to weeks per design variant in traditional CFD/FEM. It accelerates design iteration while retaining traditional solvers for validation and edge cases.
Mistral AI signed a definitive agreement this week to acquire Physics AI pioneer Emmi AI, strengthening its AI transformation services for industrial companies. Founded in Austria, Emmi AI has over 30 researchers and engineers focusing on large engineering models that replace days of computation with real-time simulation and build digital twins. Its co-founders and team will join Mistral's Science and Applied AI teams in May.
Mistral released Connectors in Studio. Built-in connectors and custom MCP are available through API/SDK for all model and agent calls, currently in public preview.
Google announced the expansion of content transparency and verification tools to Search, Gemini, Chrome, Pixel and Cloud. SynthID has watermarked over 100 billion images and videos and 60,000 years of audio. SynthID verification in the Gemini app has been used 50 million times and will reach Search and Chrome in the coming weeks.
Berkeley AI Research reviews parallel reasoning, focusing on models deciding when to decompose and parallelise independent subtasks, how many threads to generate and how to coordinate. Existing approaches including Self-consistency, Best-of-N, Tree of Thoughts, MCTS, ParaThinker, GroupThink and Hogwild! Inference mostly impose parallel structures externally rather than teaching adaptive behaviour.
Mistral AI released Workflows in public preview, positioning it as an enterprise AI orchestration layer with durable execution, observability and fault tolerance to take AI workflows from proof of concept to production.
Google DeepMind published a paper proposing Decoupled DiLoCo, splitting large-scale training into decoupled compute islands that communicate through asynchronous data streams, isolating local hardware failures.
Berkeley AI Research proposed GRASP, a gradient-based planner for learned world models. It lifts trajectories into virtual states for parallel optimisation across time, injects randomness directly into state iterations for exploration and reshapes gradients to give actions clear signals. Avoiding fragile state-input gradients in high-dimensional visual models makes long-horizon planning more practical and robust.
Mistral AI released Spaces CLI for human developers and coding agents. Commands such as spaces init and spaces dev handle scaffolding, development environments and deployment, with a corresponding flag for every interactive prompt so agents can work autonomously end to end. Each init generates context.json and AGENTS.md to provide structure and rules.
Google Research released TurboQuant, compressing the KV cache to three bits without loss of model accuracy or training and fine-tuning, alongside the QJL and PolarQuant methods.
The AIMS collaboration between Google Research and several NHS organisations published two companion studies in Nature Cancer evaluating an AI breast cancer detection system within NHS screening workflows.
Mistral AI launched Forge, a system for enterprises to build frontier-class AI models on proprietary knowledge, supporting pre-training, post-training and reinforcement learning. It handles dense and MoE architectures and, when needed, multimodal inputs, with training and governance on companies' own infrastructure.
Mistral AI investigated a vLLM memory leak in a disaggregated Prefill/Decode deployment of Mistral Medium 3.1. System memory grew linearly at 400 MB per minute, causing out-of-memory failures within hours. It occurred only on the decode side with graph compilation and KVCache transfer via NIXL. Heaptrack showed stable heap memory; the issue was outside the heap.
Mistral AI released Mistral OCR 3 with an overall 74% win rate against Mistral OCR 2 on forms, scans, complex tables and handwriting. The company says its accuracy exceeds enterprise document-processing and AI-native OCR solutions.
Mistral released the Mistral 3 family, comprising 14B, 8B and 3B small dense models and its strongest yet Mistral Large 3, a sparse MoE with 41B active and 675B total parameters. All are open-source under Apache 2.0.
Mistral AI announced a multi-year SAP partnership to integrate its models into AI Foundation and jointly develop tailored solutions for complex European industries and the public sector. It will also accelerate vision-language-action model research with Helsing for defence and security, open a German office in the coming months and significantly expand its local team.
Mistral released Mistral AI Studio, a production AI platform for enterprise teams built on three pillars: Observability, Agent Runtime and AI Registry.
Mistral fine-tuned Pixtral-12B with LoRA, substantially outperforming the untuned baseline on Aerial Image Dataset (AID) satellite classification and reducing hallucinated invalid class names. Fine-tuning uses Mistral's API or LaPlateforme UI without extensive hyperparameter tuning, with 8,000 training and 2,000 test samples.