Building with Beam

Beam at a glance:

• Architecture: sparse mixture-of-experts — 501B total parameters, 23B active per token.
• Modalities: text in, text out.
• Context: up to 1M tokens, by design.
• Weights: Apache 2.0, in BF16, FP8, and NVFP4, on Hugging Face and partner platforms.

Three ways to build

1. API — call Beam over HTTPS, or with the official OpenAI SDKs.
2. Mirror — Reflection's terminal coding agent built on Beam, or point another coding agent at Beam.
3. Deployment — run Beam on your own hardware and verify the result.

Quickstart

Create an API key at platform.reflection.ai — it is shown once, so store it as REFLECTION_API_KEY. Beam is served from an OpenAI-compatible endpoint, and a first request names the model and a list of messages:

The platform Playground lets you try prompts without code, and generates the request code for the prompt you settle on.

Capabilities

Tool calling, structured JSON outputs, and streaming are supported, along with the usual messages, roles, and parameters — plus documented errors and retries.

Reasoning effort

Reasoning is always on. The effort setting — low / medium (default) / high / xhigh / max — trades speed against depth: lower effort answers faster, higher effort thinks longer before answering.

Make Beam yours

There are two paths: adapt Beam itself (LoRA, full fine-tuning, RL), or use Beam as a teacher to train a smaller, cheaper model.

Recipes — complete, verified runs

Each recipe is a complete run, not a sketch: public or synthetic data, the actual configurations and scripts that were run, the hardware and runtime recorded, and before/after results measured the same way.

• Self-distillation — Beam fine-tunes on its own verified Triton kernels. No teacher, no human-written kernels.
• SWE agent distillation — Gemma 4 E4B learns from Beam's passing traces in synthetic software-engineering environments.
• Tool-calling reasoning — Gemma 4 E4B learns brief pre-tool thinking from thoughts Beam wrote.
• RL with Miles, NeMo RL, and SkyRL — GRPO on DAPO-Math-17K, full-parameter; SkyRL also covers LoRA.

Data

The data-synthesis folder holds the teacher data: Beam writes tool-calling thoughts for 3,941 public conversations, and builds software-engineering environments with hidden tests, keeping only passing trajectories. Self-distillation generates its own data instead.

The ladder — stop when your quality bar is met

1. Generate data with Beam as teacher — inference compute only.
2. LoRA / QLoRA — adapters, small footprint.
3. Full fine-tuning — all parameters.
4. RL post-training — verifiable reward; most capable, most compute.

Whichever rung you stop at, the pipeline stays constant: prepare data, train, evaluate against the baseline.

Compute budget

One node = 8× H200, with 141 GB HBM per GPU — roughly 1.1 TB per node.

• Data generationinference only
• QLoRA (NF4, ~250 GB)1 node · measured
• LoRA NVFP41 node
• LoRA FP81 node
• LoRA BF162 nodes
• Full fine-tuning (all 501B)3–8 nodes
• RL (all 501B)4–8 nodes

Adapter trainable parameters run ~29M–3.6B. All rows except QLoRA are estimates from memory footprint.

Folders & stacks

• data-synthesis — teacher data.
• sft — SFT and distillation.
• rl — RL with reward design and checkpoint resume.

SFT stacks: NeMo AutoModel, NeMo Megatron Bridge, Axolotl, TRL. RL stacks: Miles, NeMo RL, SkyRL. Built with NVIDIA, RadixArk, Hugging Face, Anyscale, and Axolotl.

The Open Safety Framework

The mission is to keep intelligence open and accessible: openness lets researchers inspect, stress-test, and strengthen safeguards, and lets organizations run AI under their own governance.

The Open Safety Framework (OSF) is Reflection's public commitment to identify, evaluate, and mitigate systemic risks. It covers "catastrophic risk" (US TFAIA) and "systemic risk" (EU AI Act), is informed by the White House Accord, the NIST AI RMF, and ISO 42001, and runs on standing cross-functional governance with escalation to company leaders and the Board — plus open community engagement through red teaming, evaluations, and open-weights safety systems.

Governing logic

Two risks pull in opposite directions. Concentration risk: capability held by a few leads to a durable loss of oversight. Diffusion risk: misuse from broad release. Open release reduces concentration but can raise diffusion; because concentration harm is the more durable of the two, the default is open release — with a backstop for the most severe cases (catastrophic and irreversible), judged by marginal risk, offense–defense balance, and defensive acceleration.

Eight systemic risks

• Concentration of power — tracked through ecosystem indicators.
• Established risks — biological weapons, chemical weapons, cyber offense, and loss of control.
• Research risks — radiological weapons, nuclear weapons, and harmful manipulation; these need bespoke evaluations and remain open research problems.

Tiers

Each risk is judged marginally, against comparable models. Limited means no meaningful uplift; Critical means a decisive contribution to a catastrophic outcome. At Critical, deployment mitigations apply — classifiers, system-instruction intervention, prompt block, routing, and monitoring — and release is paused until remediation plus an affirmative safety case. Without that, there is no release.

Safety margin

Before release, four uncertainties are assessed:

1. Capability uncertainty
2. Under-elicitation risk
3. Historical evaluation uncertainty
4. Mitigation-durability uncertainty

Release gate

The Model Report records the tiers and safety margin, risk estimates (with mitigations, fully elicited), the mitigations and their limits, external evaluator input, and the justification. The committee proceeds only if the result is acceptable.

Evaluations

Capability evaluations, release safety evaluations, mitigation-robustness evaluations (at Critical), third-party assurance, and threat modeling. The same bar applies to internal use.

Beam-specific

Beam's safety comes from a separate safety and alignment model (SFT + RL), merged with the capability model via multi-teacher on-policy distillation — with deliberative alignment and adversarial training against harmful compliance, excessive refusals, hallucinations, and unsafe tool use. External testers include Gray Swan, Irregular, and SecureBio, and the safety results and evaluations are slated for publication and open-sourcing.

Official documentation

This summary is a snapshot: numbers and limits in the official documentation can change (beta), and the official docs are the primary reference. Read them in full:

• developers.reflection.ai
• developers.reflection.ai/quickstart
• developers.reflection.ai/customization-overview
• reflection.ai/safety