Today for AI
HOT RADAR
llmHEAT 61.9°

AI CURATED HOT STORY · 10/16/2026

Reflection Launches 501B Open-Source MoE Model Beam for Coding and Agents

4 reports archived4 independent sourcesupdated 10/16/2026, 23:23:10
Synthesis & Latest Updates
4 Sources Cross-Validated

US AI startup Reflection has officially launched Beam, its first open-source large model. Utilizing a Mixture-of-Experts (MoE) architecture with 501B total parameters but only 23B active, Beam significantly reduces inference compute requirements while supporting a 1M token context window. Released under the permissive Apache 2.0 license, it is optimized for code generation and agentic tasks, offering various quantized versions to lower deployment barriers.

Cross-verified reports indicate strong performance in agentic workflow benchmarks, outperforming competitors like K2 Horizon with high token efficiency. vLLM provides native support from day one, achieving 708 tokens/s via DFlash2 speculative decoding. Kernel optimizations by coding agents further boosted batch decoding speed by 2.71x. Although weaker in terminal command-line operations, Beam approaches frontier closed-source intelligence levels.

This release impacts the open-source ecosystem significantly. AWS has integrated Beam into SageMaker JumpStart for managed deployment, signaling cloud providers' rapid adoption of high-performance open MoE models. For developers, Beam offers a balanced choice of long context, high speed, and low cost, particularly accelerating agent-based applications requiring multi-step tool calling and complex code maintenance.

LATEST/The vLLM team confirmed full native support for Beam, optimizing CUDA kernels via coding agents to boost prefill speed by 1.87x. Meanwhile, AWS announced Beam's availability on Amazon SageMaker JumpStart, supporting Kubernetes-native control and HyperPod Inference distributed deployment, further lowering enterprise integration barriers.

HEAT TRENDHourly heat curve

2 fully observed hours
Fewer than 3 fully observed hours; trend not plotted yet.

TIMELINECoverage timeline

Total 4 reports · Latest first
  1. AWS Machine Learning BlogT1·68 pts
    • Beam is an open-weight model from Reflection AI focused on code maintenance, tool use, and multi-step workflows.
    • The model utilizes a Mixture-of-Experts (MoE) architecture with 501B total parameters and 23B active parameters.
    • It is integrated into Amazon SageMaker JumpStart, supporting real-time endpoints and distributed deployment via HyperPod Inference.
  2. Hugging Face BlogT1·85 pts
    • Beam is Reflection's first open model, featuring 501B total parameters but only 23B active per token, reducing inference costs.
    • The model is specialized for coding and agentic use cases, supporting a 1M token context window.
    • Released under the permissive Apache 2.0 license, it includes FP8 and NVFP4 quantized versions for flexible hardware deployment.
  4. Artificial AnalysisT1·78 pts
    • Beam is Reflection's first open-weights model, utilizing an MoE architecture (501B total / 23B active).
    • Scores 48.5% on AutomationBench-AA, significantly outperforming rival K2 Horizon in agentic tasks.
    • High token efficiency: uses half the output tokens of some competitors to achieve similar intelligence levels.
    • Weakness lies in terminal command-line operations (Terminal-Bench), scoring much lower than Qwen and Muse Spark Open.
Reflection Launches 501B Open-Source MoE Model Beam for Coding and Agents | Today for AI