Overview
Beam is reflection-ai’s first open-weight frontier model — a text-only sparse MoE with 501B total / 23B active parameters, 1M token context, pretrained on 23.8T tokens. Targets coding, reasoning, and agentic workloads.
Key Specifications
- Architecture: sparse MoE, 52 layers, interleaved local+global attention
- Self-reported scores: SWE-bench Verified 80.9%, Terminal Bench v2.1 80.1%, GPQA Diamond 90.5%, AIME 2026 97.8%
- RL: 10.5K GB300 GPUs, 4 weeks, 100M+ rollouts
- License: Apache 2.0 (weights due October 2026)
Competitive Position
- Claims parity with glm-5-2 on reasoning at 3-4x lower inference compute
- Trails Kimi K3 and DeepSeek V4.1 Flash on some coding benchmarks per third-party summary (2026-10-06-reflection-beam-helpnetsecurity)
- First U.S. startup open model claiming comparable performance to leading Chinese open models (2026-10-06-reflection-beam-siliconangle)
Warning
Benchmarks unverified; weights not yet public at announcement.