CoreWeave Launches Forge: Unified Full-Lifecycle Platform for AI Models and Agents

CoreWeave has officially unveiled CoreWeave Forge, an integrated AI development layer uniting execution, observability, data curation, post-training, and evalua

tau · October 5, 2026

#CoreWeave #AI-Infrastructure #AI-Agent #LLM #Cloud

CoreWeave Launches Forge: Unified Full-Lifecycle Platform for AI Models and Agents

On September 30, 2026, CoreWeave Inc. (Nasdaq: CRWV) officially announced CoreWeave Forge at its annual AI cloud conference, Fully Connected 2026, hosted in San Francisco. The newly unveiled platform introduces a dedicated development layer engineered to unify the full lifecycle of artificial intelligence models and autonomous agents within a single connected environment.

Official launch branding and architecture visual for CoreWeave Forge AI development platform

Image source: CoreWeave

Presented before an audience of more than 4,500 enterprise customers, partners, and AI developers, CoreWeave Forge addresses widespread industry friction caused by fragmented development tooling. By unifying the continuous AI loop—Run, Observe, Curate, Improve, Evaluate, and repeat—into one coherent workflow, the platform ensures that operational signals gathered during production execution feed directly and seamlessly into subsequent model and agent refinements.

Bridging the Fragmented AI Development Loop in Production

For engineering organizations deploying foundation models and multi-agent architectures, the development lifecycle has historically been fractured across disparate tools from competing vendors.

In conventional workflows, production trace logs fail to connect automatically to subsequent training runs, completed experiment telemetry does not inform downstream evaluation suites, and manual data handoffs between disconnected systems routinely cause signal loss and project bottlenecks.

CoreWeave Forge eliminates these manual handoffs by integrating the entire iteration lifecycle into an interconnected system:

  • Unified Five-Stage Loop: CoreWeave Forge tightly connects Run (deploying models and agents to collect live signals), Observe (monitoring operational behaviors and traces), Curate (filtering and preparing actionable training data), Improve (refining model weights via post-training), and Evaluate (benchmarking releases against repeatable quality baselines) under a single account and unified navigation.
  • Open Multi-Cloud Architecture: The development layer avoids proprietary lock-in by remaining open across any foundation model, framework, or third-party cloud infrastructure an organization already operates.

Chen Goldberg, Executive Vice President of Product and Engineering at CoreWeave, highlighted the operational necessity of this integration: "As more people build with AI, they’re putting models to work with their own data and workflows. That’s where the gaps between model capability and system performance become clear. Engineering teams need to understand those gaps, identify the signals that matter, improve the next version, and measure whether the change worked under real operating conditions. All of this has to function as a single, connected system."

Integrated Subsystems: Weights & Biases Models, OpenPipe, and marimo Notebooks

CoreWeave Forge consolidates industry-standard developer tools directly into CoreWeave’s specialized GPU cloud infrastructure:

  • Weights & Biases Models & ARIA Integration: Delivers native experiment tracking, scalable hyperparameter sweeps, interactive telemetry analysis, and automated workflows. Through deep integration with CoreWeave ARIA, the platform supports autonomous research (autoresearch) loops that iteratively refine models in response to detected production signals.
  • CoreWeave Notebooks: An interactive development environment running directly within the target compute cluster. Prototypes transfer directly into distributed training, evaluation, and production serving without requiring disruptive code rewrites. Rich, interactive data visualizations are built upon the open-source marimo notebook project.
  • CoreWeave Registry: Model checkpoints and agent configuration states are versioned in open, portable formats with comprehensive lineage tracking, allowing engineering teams to branch or resume experimentation from any historical milestone.
  • CoreWeave Post-Training: Incorporating post-training capabilities from OpenPipe, Forge provides turnkey serverless mechanisms that tune model weights directly from curated production data without requiring developers to provision or operate complex training clusters.

40% Lower Costs in Serverless RL and 'RL Rollouts' Preview on Dedicated Inference

Across post-training and inference operations, CoreWeave revealed verified performance and efficiency benchmarks:

  • Serverless RL and Serverless SFT: Serverless Reinforcement Learning (Serverless RL) delivers training speeds 1.4 times faster while reducing infrastructure costs by 40 percent compared to self-managed setups. Serverless Supervised Fine-Tuning (Serverless SFT) is also available to quickly adapt open-weight models to domain-specific datasets.
  • Model Distillation (New Service): Trains smaller, faster open-weight models using the high-quality outputs of larger models for tasks already proven in production. Forge conducts head-to-head evaluation scoring between the candidate and the active model, providing the empirical proof needed to shift production traffic safely while lowering serving costs and latency.
  • Dedicated Inference and RL Rollouts Preview: For latency-critical enterprise workloads requiring reserved compute and tenant isolation, CoreWeave Dedicated Inference adds a preview capability called 'RL Rollouts'. This feature hot-loads trained checkpoints directly into live serving endpoints without taking the service down.

Enterprise Adoption with Canva and MasterClass, Plus Operational Boundaries

CoreWeave Forge debuts with enterprise production deployments already operational across major platforms.

Visual communication giant Canva and digital education provider MasterClass are actively building on CoreWeave Forge, utilizing its infrastructure to orchestrate ongoing model fine-tuning and agentic production workloads without getting bogged down in cluster maintenance.

The launch represents a strategic evolution for CoreWeave, transitioning from a specialized GPU infrastructure-as-a-service (IaaS) provider into an integrated, enterprise-grade software platform competing directly with traditional hyperscalers such as AWS, Google Cloud, and Microsoft Azure.

However, development teams evaluating Forge should note verified technical boundaries: the RL Rollouts near-zero-downtime hot-loading feature is currently in preview status and is exclusive to the Dedicated Inference tier, while standard multi-tenant Serverless Inference continues to provide catalog-based, pay-as-you-go model access.

Sources