October 2026 Summaries
14 posts from Coreweave
Filter
Month:
Year:
Post Summaries
Back to Blog
CoreWeave describes its Global AI Network as a petabit-scale backbone designed to support the large, continuous data transfers associated with AI training and inference, connecting 51 active data centers in North America and Europe with more than 1.5 GW of active power and about 4.2 GW of contracted power. The company announced a 22.5 MW renewable-energy-powered facility in Copenhagen and plans to add points of presence in the Middle East and Asia-Pacific, including the United Arab Emirates, India, Indonesia, Singapore, Australia, South Korea, and Japan, beginning in late 2026. Its network architecture uses multiple 400 Gbps connections at each point of presence, hardware-based end-to-end MACsec encryption, and redundant terrestrial and subsea routes intended to preserve performance, security, and availability during disruptions. CoreWeave says the backbone integrates with its data-center network fabric, bare-metal infrastructure, and object storage technology to reduce bottlenecks in moving datasets and model artifacts to GPUs, with the aim of increasing model utilization, goodput, and time to results.
Oct 27, 2026
1,294 words in the original blog post.
CoreWeave, NVIDIA, and You.com completed a three-week effort to post-train NVIDIA’s open-weight Nemotron 3.5 Lightning model for agentic web search, raising BrowseComp accuracy from 36.97% to 45.45% while reducing average tool calls by 31.55%, from about 168 to 115 per task. You.com created web-grounded multi-turn question-answer data, filtered tasks by difficulty using repeated model attempts, and used 4,932 supervised fine-tuning traces followed by 1,480 GRPO reinforcement-learning tasks with binary rewards. CoreWeave’s managed RL Rollouts infrastructure and NVIDIA’s Dynamo, ModelExpress, and vLLM technologies sought to minimize a major RL training bottleneck by replacing repeated inference-server restarts with in-place “Hot-Load” weight updates, using compressed checkpoint deltas and GPU-to-GPU transfers; reported tests showed the fastest delta update was roughly 15 times faster than an average cold-start redeploy. The collaboration attributes improved search behavior to both targeted data and efficient training infrastructure, with the optimized model exhibiting less repetitive searching, better use of retrieved evidence, and more confidence in verified answers. The partners describe the approach as repeatable for other post-training domains and plan to expand training data, incorporate efficiency-aware rewards and agent budgets, and make the You.com search integration available through NVIDIA NeMo Gym.
Oct 22, 2026
1,696 words in the original blog post.
CoreWeave has announced its Zero Trust AI Framework, a customer-controlled security model for sensitive AI workloads that aims to reduce reliance on cloud-provider policies by using technical controls across data, models, infrastructure, and runtime environments. The framework includes customer-managed Remote Key Encryption, customer-authorized support access through Support Access Management, hardware-rooted remote attestation, confidential computing planned for NVIDIA Blackwell and Vera Rubin platforms, a commitment not to use customer prompts, training data, weights, or outputs to improve CoreWeave’s own models, and MACsec encryption for data moving across its network backbone. CoreWeave says these measures allow customers to retain control of encryption keys and administrative access, verify infrastructure before releasing secrets, protect data and code while in use, and secure transfers between data centers. The company positions the framework as particularly relevant for enterprises handling regulated data or proprietary intellectual property as AI agents and reinforcement-learning workflows expand the number of systems and artifacts that may process sensitive information.
Oct 21, 2026
1,830 words in the original blog post.
NVIDIA HGX Rubin NVL8, scheduled to arrive on CoreWeave Cloud this quarter, is an eight-GPU system based on the Vera Rubin architecture designed for agentic AI, large-scale analytics, and high-performance computing workloads constrained by memory capacity, bandwidth, and GPU communication latency. It combines 2 TB of HBM4 memory, 130 TB/s aggregate memory bandwidth, sixth-generation NVLink with 24 TB/s within-node bandwidth, ConnectX-9 networking, and BlueField-4 infrastructure offload; NVIDIA projects up to tenfold higher token throughput and comparable DeepSeek-R1 training performance with four times fewer GPUs than HGX B200 in cited tests. CoreWeave plans to support the platform through liquid-cooled infrastructure monitoring and control tools, managed Kubernetes, a Slurm-Kubernetes scheduling environment, and direct engineering support. Suggested use cases include serving large-context, multi-agent AI systems, accelerating distributed data pipelines and analytics, and running double-precision scientific simulations or hybrid HPC-AI workflows.
Oct 21, 2026
1,048 words in the original blog post.
Reflection’s open-weight Beam model is now available through CoreWeave Forge Serverless Inference via an OpenAI-compatible, per-token API, allowing developers to test coding and long-horizon agentic workloads without managing GPU infrastructure. Beam is an Apache 2.0-licensed mixture-of-experts model with 501 billion total parameters, 23 billion active parameters per token, and an approximately one-million-token context window, designed for source code, tool calls, structured outputs, and multi-step software engineering tasks. Reflection used Weights & Biases during training, while CoreWeave and Axolotl validated supervised fine-tuning examples on GPU kernel generation and medical question answering, including reported HealthBench gains for a fine-tuned medical model. Reflection’s published benchmark figures and CoreWeave’s production evaluations indicate comparable quality results, while CoreWeave reports low first-token latency and high token-generation speeds on NVIDIA GB300 systems across chat, agentic, and long-generation workloads. The company advises teams to assess complete task success, tool reliability, elapsed workflow time, and cost per successful task, using tracing and evaluation tools to determine whether managed inference, fine-tuning, or reinforcement learning could improve results for their applications.
Oct 16, 2026
1,309 words in the original blog post.
AI systems expand traditional security reviews by introducing rapidly changing assets such as model weights, training data, prompts, traces, evaluations, and agent workflows that move across training, inference, and production environments. The passage argues that security must shift from static environment-based approval toward runtime-verifiable controls that remain attached to work as it evolves, including identity management, key authority, data-use restrictions, telemetry, revocation, and audit evidence. It describes this continuous development cycle as the “AI loop,” in which production behavior informs data curation, model improvement, evaluation, and redeployment, creating risks of control drift when governance no longer reflects the active system. CoreWeave presents Remote Key Encryption as a way for customers to retain control of encryption keys while using its AI infrastructure, while distinguishing protection at rest from future needs around confidential computing and attestation for data in use. The passage recommends integrating AI environments with existing enterprise security operations and evaluating platforms based on enforceable sandboxing, identity and key controls, revocation speed, SOC telemetry, documented capabilities, and verifiable protection across the AI lifecycle.
Oct 15, 2026
1,794 words in the original blog post.
AI data center anomaly detection can combine two distinct uses of agents: offline agents that build and refine deterministic machine-learning pipelines, and live agents that investigate the relatively few anomalies those pipelines flag. The deterministic detector processes millions of telemetry signals consistently using versioned features, models, thresholds, and post-processing, while agents are used during development within a controlled evaluation harness and under engineer review. For flagged events, a live reasoning agent can gather telemetry, node information, incident history, and runbooks to assess likely causes and recommend actions, but its claims must be supported by tool-derived evidence and final operational decisions remain with human operators. This arrangement aims to detect emerging failures before they affect customers while avoiding the cost, inconsistency, and alert fatigue of applying language-model reasoning to every signal. Over time, recurring patterns identified during investigation can be proposed for conversion into deterministic rules, making the system more predictable and potentially less expensive as it scales. The approach depends on extensive historical data, experienced operators, and engineering infrastructure, and may also apply to other sensor-rich physical systems such as batteries, robotics, and vehicles.
Oct 15, 2026
2,123 words in the original blog post.
CoreWeave describes Direct-to-Expert as a technical support and engineering model for AI training and inference workloads, designed to connect customers directly with cross-functional engineers who remain involved from initial workload design through production incidents. The company says its ARENA evaluation environment allows teams to test real models and pipelines on production infrastructure, assess performance and costs, and tailor cluster, storage, networking, and scaling designs before deployment. It emphasizes that AI infrastructure failures can span schedulers, storage, and networking, and argues that rapid access to engineers familiar with a customer’s environment can reduce debugging time and operational disruption. CoreWeave also highlights Mission Control for infrastructure monitoring and automated fault repair, along with specialized field engineering for physical AI, robotics, and simulation workloads. Customer comments from Cohere and IBM are presented as examples of the approach’s value, while the company invites prospective users to request an ARENA Pass for a guided evaluation.
Oct 07, 2026
1,215 words in the original blog post.
CoreWeave argues that model development should continue after deployment, using production traffic to identify failures, collect training signals, and improve quality or cost efficiency through distillation, supervised fine-tuning, or reinforcement learning. Announced at Fully Connected 2026, CoreWeave Forge combines Model Distillation, a preview programmable Serverless Training API, and RL Rollouts on CoreWeave Inference to connect production observations with training and redeployment. Model Distillation captures inference data through Agent Lens, supports relabeling and experimentation, and can help smaller models match desired behavior, as illustrated by Method’s updated Llama 3.1 8B model winning 57.5% of comparisons against its prior production version. The programmable API allows teams to write customized Python training loops that combine approaches such as RL and supervised learning while CoreWeave manages underlying infrastructure, while RL Rollouts supports rapid checkpoint hot-loading, training-specific metadata, and continuously synchronized serving policies. CoreWeave, NVIDIA, and you.com reported that this approach improved a web-search model’s BrowseComp accuracy from 36.97% to 45.45% while reducing tool calls by 30.24%, with checkpoint updates reportedly about 15 times faster than conventional redeployment.
Oct 05, 2026
1,636 words in the original blog post.
CoreWeave announced Forge at Fully Connected 2026 as an integrated, open AI development environment intended to connect the continuous lifecycle of deploying, observing, curating data, improving, and evaluating AI models and agents. The platform addresses operational fragmentation between research, engineering, and production teams by linking production traces, user feedback, datasets, evaluations, model versions, and deployments in a shared system of record. Forge includes Agent Lens for agent observability and failure analysis, ARIA for AI-assisted research and coding, Sandboxes for isolated execution, Model Distillation, managed Notebooks, inference services, post-training capabilities, and Registry asset management. CoreWeave also expanded Dedicated Inference with RL Rollouts to support rapidly updating reinforcement-learning policies, and described partnerships with technology, data, and web-search providers. Designed to work across models, frameworks, clouds, and on-premises environments, Forge aims to let teams retain portable data and existing tools while shortening the path from production issues to validated improvements.
Oct 02, 2026
2,060 words in the original blog post.
CoreWeave Registry is presented as a centralized system of record for managing AI assets—including models, agents, datasets, prompts, and evaluation results—through development, approval, deployment, and monitoring. It addresses operational challenges such as identifying the correct version, tracing an asset’s data and parameters, documenting approvals, reproducing results, and safely updating production systems. Its core capabilities include immutable versioning for reproducibility, movable aliases such as staging or production for controlled release workflows, detailed lineage for tracking inputs and origins, including imported Hugging Face models, and role-based permissions for governance. Registry connects experimentation with automated CI/CD processes by enabling evaluation in CoreWeave Sandboxes, policy checks, reviews, alias assignment, and deployment triggers. It also supports organization-wide discovery and reuse of governed assets while maintaining access controls, and it extends lifecycle management beyond individual models to the interconnected components of complete AI applications.
Oct 02, 2026
1,423 words in the original blog post.
CoreWeave’s first Fully Connected conference in San Francisco brought together more than 3,000 AI practitioners to discuss deploying AI systems in production, with customers including Caterpillar, Capital One, and Cognition describing enterprise applications. The company argued that growing AI workloads require infrastructure designed specifically for AI rather than legacy computing systems, and presented itself as a cloud platform supporting that need. Announcements included CoreWeave Forge, a platform for building, running, and improving AI models and agents; availability of NVIDIA Vera Rubin NVL72 systems, which Cognition is using for production workloads; and NVIDIA Vera CPUs designed to support AI-agent tasks such as sandboxes and tool calls. CoreWeave also highlighted its partnership with home-care provider Ennoble Care for clinical AI workloads and introduced the CoreWeave Partner Network, featuring tested integrations from companies including CrowdStrike, VAST Data, ClickHouse, and Reflection.
Oct 02, 2026
723 words in the original blog post.
CoreWeave has introduced Notebooks in public preview within Forge, providing project-connected Python and SQL environments for investigating model behavior beyond aggregate dashboard metrics. Powered by the open-source reactive notebook platform marimo, the notebooks automatically update dependent outputs when code or variables change, helping keep analyses, program state, and visualizations aligned during debugging. An example involving a driving model shows how slicing data by weather, time of day, and nearby red objects can reveal a failure mode—confusing traffic lights with red buildings or trucks in heavy nighttime rain—that average loss metrics concealed. Notebooks are stored and shared within Forge projects, allowing teammates to discover, reuse, rerun, and turn prior investigations into regression checks across new experiment data. CoreWeave plans to add persistent Workspace panels, deployable data apps, expanded compute for model training, and ARIA-assisted notebook authoring to support users who need custom analyses without extensive familiarity with data-science libraries.
Oct 01, 2026
1,026 words in the original blog post.
CoreWeave has announced the general availability of ARIA, an AI Research and Iteration Agent integrated into its Forge platform to help teams analyze, improve, and evaluate AI models and agents through iterative experimentation. ARIA can review experiment configurations, metrics, artifacts, training behavior, and connected code repositories; generate visual reports and interactive charts; identify performance drivers; recommend follow-up experiments; draft code changes in GitHub branches; and launch approved jobs on users’ infrastructure. It supports workflows ranging from prompt and agent evaluation using production traffic to model pre-training, fine-tuning, supervised fine-tuning, and reinforcement learning, with users able to retain approval control or automate the full cycle. Running in the cloud, ARIA can continue analyses and respond to event-based automations after a user disconnects, while its project-scoped memory preserves shared team context, decisions, and findings. The tool also integrates with GitHub, Hugging Face, and Kaggle, and is accessible through the Forge dashboard, ARIA home page, and an iOS application.
Oct 01, 2026
797 words in the original blog post.