CoreWeave Launches Forge, Bundling Weights & Biases With Serverless Training and Inference

CoreWeave just rolled out Forge, a platform that stitches together Weights & Biases experiment tracking, agent evaluation tooling, and serverless infrastructure for inference, post-training, and sandboxed compute, all running on CoreWeave's own GPU backend. If you've used Weights & Biases on their Multi-tenant Cloud before, the migration is painless: your projects, runs, artifacts, and account carry over automatically.

The pitch here is consolidation. Instead of bolting together separate tools for tracking experiments, evaluating agents, fine-tuning models, and serving inference, Forge tries to put all of it under one roof with shared infrastructure underneath. For teams tired of context-switching between five different dashboards to answer "why did this run regress," that's a meaningful quality-of-life improvement, assuming the integration actually holds up in practice.

What's Actually in the Box

Forge splits into two broad categories: tools for tracking and improving your work, and infrastructure for running it.

On the tracking and evaluation side:

• Weights & Biases (W&B Models) handles the fundamentals: experiment tracking, hyperparameter sweeps, artifacts, reports, and automations. This is the part most ML engineers already know. W&B Weave sits alongside it specifically for evaluating LLM and agent applications, which is a distinct problem from tracking a standard training run since you're often grading output quality rather than a clean loss curve.
• ARIA is an AI research agent that analyzes your experiments, explains results in plain language, and builds visualizations and reports on its own. Think of it as an assistant layer sitting on top of your experiment data rather than another dashboard you have to manually query.
• Agent Lens focuses on observability and improvement specifically for agents, a category that's gotten messier as more teams ship multi-step, tool-using LLM systems that are harder to debug than a single forward pass.
• Notebooks (currently in preview) are reactive Python notebooks built for exploring data, prototyping model workflows, and digging into runs and artifacts without leaving the platform.
• Registry gives you a versioned hub for models and datasets with lineage tracking and a defined handoff point into deployment, which matters once you've got more than one model in flight and need to know which dataset version actually produced which checkpoint.

On the infrastructure side, Forge offers three serverless products:

• Post-Training lets you fine-tune and post-train LLMs using Serverless RL (reinforcement learning) and SFT (supervised fine-tuning), plus a distillation path for building specialized smaller models from real application traffic rather than synthetic data.
• Serverless Inference exposes open-source foundation models through an OpenAI-compatible API, already integrated with Weave for evaluation. The OpenAI-compatible interface is the practical bit here: it means you can likely swap this in without rewriting your client code.
• Serverless Sandbox gives you on-demand, isolated compute environments that you spin up and tear down directly from Python, useful for anything that needs a clean, disposable execution context like running untrusted or agent-generated code.

CoreWeave is also pointing people toward a Builder Resource Center with demos, guides, code samples, and events for working across these products, which suggests they're expecting a ramp-up period as teams figure out how the pieces fit together.

Key Takeaways

The interesting architectural bet in Forge is treating post-training and inference as serverless from the jump. That's not a new idea in the broader industry, but tying it directly to an experiment-tracking and evaluation layer that was previously a standalone product (W&B) is a notable move. It suggests CoreWeave wants to own more of the ML lifecycle on its infrastructure rather than just being the GPU layer underneath someone else's stack.

Serverless RL specifically is worth watching. RL-based post-training for LLMs has historically required a decent amount of custom infrastructure, reward modeling pipelines, rollout generation, and the compute orchestration to make them efficient. Wrapping that into a serverless offering alongside more standard SFT is a sign this workflow is maturing from research-lab bespoke setups into something closer to a managed product.

The agent evaluation angle, split across Weave and Agent Lens, also reflects where a lot of practitioner pain currently sits. Evaluating a single model's output against a benchmark is relatively well understood at this point. Evaluating a multi-step agent that calls tools, makes decisions, and chains actions together is a much messier problem, and tooling for it is still catching up to the pace at which people are actually deploying agents in production.

Whether Forge delivers on the "no infrastructure management" promise will come down to execution details that aren't fully spelled out yet: latency on the serverless inference endpoints, pricing for sustained post-training workloads, and how well ARIA's automated analysis holds up against hand-rolled debugging when a run goes sideways. For teams already in the Weights & Biases ecosystem, though, the migration path looks low-friction, and that alone lowers the bar for trying it.