3AM Marketer
AI

Reflection Beam 501B Specs, Architecture, and APIs

Reflection AI unveiled Beam, a 501-billion-parameter sparse MoE model activating 23 billion parameters per token. Here are its architecture details, verified benchmark scores, and API limits.

23B active
SWEBench Verified80.9
SWE Bench Pro v2-Hard77.2
SWE Bench Pro v165.5
SWEBench Multilingual78.0
Score, from the entry's own table. 6 sources, checked 16 Oct 2026. Full table

Reflection AI announced Beam on October 5, 2026, introducing a sparse Mixture-of-Experts architecture with 501 billion total parameters that activates 23 billion parameters per generated token. Pretrained on 23.8 trillion tokens, the model delivers reasoning scores comparable to GLM-5.2, which costs $1.40 input and $4.40 output per million tokens in our API pricing data, while requiring 3 to 4 times less inference compute.123

The 52-layer architecture and cluster training dynamics

Beam organizes its weights across 52 layers using a 3:1 interleaved global and sliding-window attention mechanism. Load balancing builds on auxiliary-loss-free balancing adapted from DeepSeek-V3, incorporating a cosine decay for expert-bias updates. This routing kept the busiest expert capped at 1.04x average load at the conclusion of pretraining.456

Pretraining finished in under 4 weeks across 6,144 NVIDIA GB300 NVL72 GPUs. The cluster maintained 92.3% useful goodput toward the end of the run with 9 checkpoint rewinds, achieving an estimated ~12% BF16 model flops utilization (MFU). Layer stability across all 52 layers was maintained using depth-based scaling, SandwichNorm, attention gating, and FP32 residual accumulation.456

Reinforcement learning at scale across 1.3 billion sandboxes

The post-training reinforcement learning run spanned 4 weeks on 10.5K NVIDIA GB300 GPUs, generating over 100 million rollouts up to a maximum context length of 256K tokens. Training and evaluation environments utilized approximately 1.3 billion execution sandboxes, running an average of 110K concurrent rollouts with peak capacity reaching 170K sandboxes.14

The asynchronous training pipeline demonstrated resilience against distributed infrastructure churn. Numerics remained stable despite one-day asynchronous staleness where rollout workers ran up to 107 weight versions behind the active policy. Weight propagation from the trainer to inference servers achieved a median latency of roughly 12 seconds, allowing the job to absorb 71 inference incidents without interrupting training.14

Software engineering and coding evaluation scores

Beam targets software engineering workflows, matching or exceeding state-of-the-art specialist tools across standard repository benchmarks for teams building open source AI coding agents.1

BenchmarkEvaluation SplitScore
SWEBench VerifiedStandard test set80.9
SWE Bench Pro v2-HardHard subset77.2
SWE Bench Pro v1Full split65.5
SWEBench MultilingualPolyglot repository split78.0
Terminal Bench v2.1CLI navigation & bash tasks80.1
DeepSWE v1.1Autonomous coding pipeline44.4
SciCodeScientific Python problems49.7
SWE Atlas Codebase QnACode repository question answering34.6
  • SWEBench Verified80.9
  • SWE Bench Pro v2-Hard77.2
  • SWE Bench Pro v165.5
  • SWEBench Multilingual78
  • Terminal Bench v2.180.1
  • DeepSWE v1.144.4
  • SciCode49.7
  • SWE Atlas Codebase QnA34.6
SWE-bench and terminal evaluation scores for Beam, October 2026.1

Reasoning, agentic, and long-context benchmark results

On broad technical evaluation suites, Beam recorded 97.8 on AIME 2026, 90.5 on GPQA Diamond, and 36.2 on Humanity's Last Exam (HLE) without external tool use. In agentic execution environments, it logged 78.7 on MCP Atlas, 38.0 on tau3 banking, 37.0 on the public AutomationBench suite, and 16.3 on CriPT AA. Structured spatial reasoning tests yielded a 95.5% accuracy rate classifying 16,200 points on a 180x90 global land-water grid.14

  • AIME 202697.8
  • GPQA Diamond90.5
  • HLE no tools36.2
  • MCP Atlas78.7
  • tau3 banking38
  • AutomationBench37
  • CriPT AA16.3
Reasoning and agentic benchmark performance for Beam, October 2026.1

For retrieval and extended reasoning, Beam reached 77.4 on BrowseComp and 80.1 on DeepSearchQA when paired with context management. On extended context benchmarks, the model posted 65.5 on LongBench v2, 79.3 on AA-LCR, 79.7 on IFBench, and 13.0 on the public split of the AA Omniscience index.1

API limits, reasoning effort settings, and self-hosted weights

While midtraining extended Beam's internal context window to 1 million tokens, the current beta API imposes a total request cap of 262,144 tokens (input plus output combined), with generation output capped at 131,072 tokens. The API provides five distinct reasoning effort settings: low, medium (the default), high, xhigh, and max. Reasoning tokens cannot be switched off completely.4

For teams reviewing self-hosted AI agents, open weights across FP8 and NVFP4 formats are scheduled for public release in late October 2026.6