Reflection AI has introduced Beam, its inaugural open-weight Mixture-of-Experts (MoE) model, specifically engineered for advanced coding and agentic workloads. This sophisticated model features 501 billion total parameters, with 23 billion active parameters processed per token. Beam is strategically positioned to compete with more extensive open models, such as GLM 5.2, by delivering comparable reasoning performance while requiring significantly less inference compute—reportedly three to four times less.
Accessibility and Strategic Focus
Beam is currently undergoing final safety evaluations and is not yet available for self-hosting. Early access is being facilitated through a waitlist on the Reflection platform. The company envisions Beam as a pivotal contribution to the Western open-weight AI ecosystem, emphasizing efficiency as its core competitive advantage. While acknowledging that some models like Kimi K3 may offer higher raw capabilities, Beam's design prioritizes inference efficiency, making it an attractive option for optimizing computational resources in enterprise settings.
A notable user-facing feature is the 'reasoning effort' parameter, allowing teams to adjust the model's output from concise answers to more elaborate reasoning, tailored to task complexity and budget constraints.
Underpinning Technology: Pretraining and Architecture
Beam's development began with pretraining on an immense dataset of 23.8 trillion tokens, sourced from public web data and proprietary licensed collections. Reflection AI's rigorous curation process involved discarding approximately 95% of raw internet tokens, while carefully preserving 1.8 trillion high-quality tokens often overlooked by conventional filters.
The model's architecture skillfully interleaves local and global attention mechanisms with fine-grained routed experts. Its advanced load balancing system, inspired by DeepSeek-V3's auxiliary-loss-free method and enhanced with cosine decay for expert-bias updates, maintained exceptional workload distribution, with the busiest expert reaching just 1.04 times the average load. The pretraining phase was completed in under four weeks on 6,144 NVIDIA GB300 NVL72 GPUs, achieving a high goodput of 92.3%. The effective context window was extended to 1 million tokens mid-training.
Scalable Reinforcement Learning
Reinforcement Learning (RL) formed a crucial scaling axis in Beam's development, involving an intensive four-week phase on 10,500 NVIDIA GB300 GPUs. Over 100 million rollouts were generated, utilizing a maximum context of 256,000 tokens. Training and grading encompassed an estimated 1.3 billion sandboxes across nearly one million distinct coding, agentic, and STEM environments.
Reflection AI employed fully asynchronous policy gradients, ensuring stable learning even with one-day staleness across 107 weight versions. Infrastructure metrics highlight the system's robustness, sustaining 110,000 concurrent rollouts on average and integrating new weights into the inference fleet in a median of 12 seconds. This phase also introduced a controllable length penalty to encourage more concise solutions and saw an unexpected improvement in browsing capabilities without explicit browsing tasks, suggesting successful transfer learning.
Safety, Benchmarks, and Market Position
Safety and alignment were paramount, with a dedicated safety teacher trained from the pretrained checkpoint and merged with the RL teacher using multi-teacher on-policy distillation. The full safety evaluation results are expected in an upcoming technical report.
In reported benchmarks, Beam demonstrated strong performance: an 80.9 on SWE-bench Verified (compared to Nemotron 3 Ultra's 70.7) and an 80.1 on Terminal Bench v2.1. While trailing DeepSeek V4.1 Flash (90.6) and Kimi K3 (88.3) on Terminal Bench, its scores remain highly competitive. These figures, as cited by Reflection AI, draw upon data from Artificial Analysis and DataCurve.
Beam stands out among its peers due to its comparatively smaller total parameter count and 23 billion active parameters, reinforcing its efficiency-driven design. Reflection AI intends to release the model weights under the permissive Apache 2.0 license later this month, aligning with industry standards for open-source AI accessibility.
This article is a rewritten summary based on publicly available reporting. For the original story, visit the source.
Source: MarkTechPost
