Volume 343, Issue 2 · 6 articles from 2 issues
CoreWeave explains how Model Distillation, RL Rollouts and a training API in CoreWeave Forge turn production inference signal into better models.
CoreWeave argues that post-training should begin once a model ships, using production traces to improve model quality and economics.
At Fully Connected 2026 it introduced CoreWeave Model Distillation, CoreWeave RL Rollouts on CoreWeave Inference and a programmable training API in preview, all brought together in CoreWeave Forge.
The post advises choosing distillation, supervised fine-tuning or reinforcement learning based on the failure mode and available signal, citing a Method example where a retrained Llama 3.1 8B model won 57.5% of head-to-head evaluations against the production version.
CoreWeave and You.com explain how faster data movement and weight loading keep GPUs busy during continuous AI agent post-training.
At theCUBE's Fully Connected event, CoreWeave SVP Corey Sanders and You.com CPO Saurabh Sharma discussed keeping GPU utilization high as enterprises continually refine AI agents.
CoreWeave Forge connects model deployment, evaluation and improvement, and its RL Rollouts capability, now in preview, supports repeated training cycles with hot-start weight synchronization from nearby peers.
Working with Nvidia, the companies post-trained Nemotron 3.5 Lightning in eight hours using You.com's web search tools.
CoreWeave details Direct-to-Expert, a support model that puts customers in a shared channel with the engineers who built its cloud.
CoreWeave describes Direct-to-Expert, a support approach with no non-technical layer between customers and the engineers who can resolve their issues.
Customers post problems in a shared channel, a human responds within minutes, and the engineer who owns the affected layer stays on the issue through resolution.
Engagements start before production in CoreWeave ARENA, a structured lab for running real models and pipelines on production AI infrastructure, and quotes from Cohere and IBM cite faster debugging on frontier training runs and help designing bursty sandbox capacity.
CoreWeave explains using AI agents offline to build a deterministic data center anomaly detector, and live only to reason about flags.
CoreWeave describes using AI agents to find data center nodes likely to interrupt training runs by monitoring power, inlet temperature, coolant flow and fan speed.
Rather than pointing agents at everything, which makes systems slow and inconsistent, it split their work so that offline, agents build a deterministic anomaly detector from conventional ML and data processing components, and live, they reason only about what that detector flags.
CoreWeave says this split was the most useful lesson from the project.
Reflection's first open-weight model, Beam, is available on day zero via Serverless Inference in CoreWeave Forge.
Reflection's first open-weight model, Beam, is available on day zero through Serverless Inference in CoreWeave Forge, callable via an OpenAI-compatible API with pay-per-token pricing.
Beam was trained using Weights & Biases Models within Forge, and CoreWeave validated supervised fine-tuning with Axolotl on its infrastructure.
The Apache 2.0-licensed mixture-of-experts model has 501 billion total and 23 billion active parameters and targets software engineering and long-horizon, tool-driven tasks.
CoreWeave argues AI adds dynamic assets like weights, checkpoints and agent paths that static CISO security reviews miss.
CoreWeave says AI adds a new class of sensitive assets with a dynamic lifecycle, including model weights, training data, checkpoints, embeddings, prompts, traces and inference paths that span multiple environments.
Agents further change the control model because they call tools, retrieve context and create action paths a review may not have anticipated.
As systems become more agentic, trust depends less on a single boundary and more on what can be enforced, observed, revoked and verified at runtime, changing how CISOs review AI.