Reflection's open-weight Beam model launches on CoreWeave Serverless Inference

A mixture-of-experts model for coding and agents

Beam is a mixture-of-experts model with 501 billion total parameters and 23 billion active parameters per token. CoreWeave explains that each token activates only a fraction of the model, which reduces computation compared with activating every parameter. Serving still has to account for the full set of weights, so the active parameter count alone does not determine memory requirements or latency.

Reflection built the release for text-native workloads such as source code, tool calls, logs, documents and structured outputs. The two target workloads are software engineering and long-horizon agentic tasks, including resolving GitHub issues, maintaining code and completing workflows through repeated model calls and tool interactions.

Beam has an approximately one-million-token context window. CoreWeave cautions that the context length exposed by an endpoint depends on the provider's serving configuration, so developers should check the supported limit when designing workflows.

The model ships under Apache 2.0 without custom license terms. Developers and researchers can inspect the weights, fine-tune the model and choose where to deploy it. CoreWeave says the release adds a new model to the Western open-weight ecosystem.

Training and fine-tuning on CoreWeave

According to CoreWeave, Reflection used Weights & Biases Models within Forge during Beam's training. CoreWeave also worked with Axolotl to validate a supervised fine-tuning (SFT) use case on its infrastructure. SFT trains a model on examples of desired behavior, which can capture things like repository conventions or expected tool-call formats.

In the first validated run, Axolotl fine-tuned Beam-1.0-501B-A23B on GPU kernel generation. The training data was about 84 million tokens of PyTorch-to-Triton and CUDA examples from the open KernelBook and AI CUDA Engineer datasets. The run used LoRA adapters with expert parallelism across 16 NVIDIA B200 GPUs on two CoreWeave nodes and finished in 73 minutes.

The second use case was medical question answering. Axolotl trained LoRA adapters directly on Beam's NVFP4 checkpoint on a single B200 node, using APO-Zero preference optimization on rubric-scored examples of Beam's own answers to patient questions. Because the adapters are trained to survive requantization, they fold back into an NVFP4 checkpoint that serves on Blackwell in the same format as the base model.

On HealthBench, the merged model's score rose from 49.0% to 52.7% across all 5,000 conversations, graded with GPT-6 Luna Decisions. CoreWeave says the improvement held up after length adjustment and under every grader tested. The company stresses that these were demo runs and do not represent Beam's full potential with fine-tuning.

CoreWeave is exploring broader support for fine-tuning and reinforcement learning with Beam, including further work with Axolotl on online RL. Enterprises that want to post-train Beam on their own data can contact the company.

Evaluating Beam through Serverless Inference

Serverless Inference offers an OpenAI-compatible endpoint with per-token billing, which CoreWeave says simplifies integration with clients and frameworks that support that interface. The company notes that model behavior still needs evaluation, since prompt conventions, tool calling, output formatting and failure recovery can differ between models.

CoreWeave Agent Lens provides tracing, evaluations and observability. For agents, the company argues the useful unit of analysis is the complete task, including model calls, tool responses, retries and the final output. It recommends measuring four things:

• Task success: whether the model produces a correct patch or completes the workflow, checked against tests or explicit acceptance criteria.
• Tool-use reliability: whether it picks the right tool, generates valid arguments and recovers from errors.
• Completion time: elapsed time across all model calls and tool interactions.
• Cost per successful task: token spend across the workflow, including failed attempts and retries.

These measurements help separate a model that does well on isolated prompts from one that works reliably inside an application.

What it means for developers

Teams can start with the managed endpoint, test Beam on their own coding and agentic workloads, and then decide where fine-tuning might help. CoreWeave offers a free trial of Forge Pro, and teams interested in fine-tuning on their own data or in reinforcement learning can contact the company about an evaluation through CoreWeave ARENA or on their own CoreWeave cluster.