Beam: Reflection's 501B Open-Weight Model

Announcement: Reflection is proud to unveil Beam, our inaugural open-weight model, designed specifically to push the boundaries of reasoning, coding, and agentic autonomy.

Published on October 5, 2026 | $\approx 8$ min read

🚀 Overview of Beam

Beam is engineered as a sparse Mixture-of-Experts (MoE) architecture. While it boasts a massive total parameter count, it maintains high efficiency by only activating a fraction of those parameters during inference.

Technical Specifications

The model's strength is derived from a dual-pronged investment: massive-scale pretraining on diverse web and licensed data, and a groundbreaking approach to Reinforcement Learning (RL).

---

🛠️ The Engineering Behind the Model

To achieve frontier-level performance, Reflection developed a specialized infrastructure capable of sustaining extreme compute loads.

The RL Pipeline

To facilitate this, the team utilized a staggering amount of hardware and data:

• Compute: 10.5K NVIDIA GB300 GPUs.
• Duration: 4 weeks of continuous training.
• Scale: Over 100 million rollouts.
• Context: Maximum window of 256K tokens.
• Infrastructure: $\approx 1.3$ billion sandboxes and 1 million specialized STEM/coding environments.

Reflection believes this represents one of the smallest one of the largest scale RL runs ever executed by an open lab.

---

📊 Model Capabilities & Benchmarks

Beam was specifically optimized for agentic performance. While some models like Kimi K3 still lead in raw capability, Beam provides a superior balance of power and inference efficiency.

Performance Comparison Tables

1. Agentic Coding & Terminal Use

2. Reasoning & STEM

3. Tool Calling & Search

4. General Capabilities

---

⚡ Efficiency and Compute

The defining characteristic of Beam is its inference efficiency. It provides a high level of "intelligence per token," making it an ideal workhorse for enterprise-grade agentic tasks.

Compute Analysis

When compared to GLM-5.2, Beam achieves similar reasoning scores while utilizing only 3–4$\times$ less inference compute. This gap widens significantly when compared to 2T+ parameter models like Qwen 3.8-Max.

To estimate the forward-pass compute, the following formula was used:

$\text{FLOPs} \approx 2 \times \text{active parameter count} \times \text{mean generated tokens per attempt}$

# Example Logic for Compute Estimation
def estimate_flops(active_params, mean_tokens):
    # Each multiply-add is counted as two operations
    return 2 * active_params * mean_tokens

Note: These estimates focus on the generation phase and exclude prompt prefill, attention overhead, and serving costs.

📈 Scaling the RL Axis

Reflection treated RL as a primary scaling axis. By increasing RL compute, the team observed a consistent rise in performance across Terminal-Bench 2.1, HLE, and DeepSWE, with no evidence of a plateau.

For context, the scale of rollouts used for training was:

• MiMo: 753K rollouts
• Inkling: 30M rollouts
• Beam: 100M+ rollouts

As the scale increased, the team noted that policy staleness became a major source of insta...