501B Parameters, Only 23B Active: Reflection's Beam Puts Sparse Efficiency at the Center of Inference

Forum topic · QianXun · 2026-10-06

Summary

On October 5, Reflection published technical details of Beam, a sparse mixture-of-experts model with 501 billion total parameters but only about 23 billion activated per token—roughly 4.6% of the model—making the effective compute bill an inference-side story rather than a storage one. Reflection itself concedes that Kimi K3 still leads on raw capability: Beam trails Kimi K3 on all four shared benchmarks, including DeepSWE v1.1 (44.4 vs 68.0) and SWE Atlas repository QA (34.6 vs 68.0). Beam's pitch is efficiency: it reportedly matches higher scores on advanced reasoning benchmarks with 3–4x less inference FLOPs than GLM-5.2, under a self-declared approximate formula excluding prefill and serving overhead. Pretraining used 23.8 trillion tokens on 6,144 GB300 GPUs in under four weeks, discarding ~95% of raw web tokens; a four-week RL stage ran 100M+ rollouts across roughly 1 million sandboxed coding/STEM environments with 1.3 billion sandbox calls. Weights are planned under Apache 2.0, though not yet released at publication. This article translates and examines the full disclosure, flagging which figures are company-reported versus community-derived estimates.

Key points

Title number vs. active compute: Beam has 501B total parameters but activates only 23B per token (~4.6%). The total determines disk space; the active count determines your inference bill. Beam is a sparse MoE model where each token routes to a small subset of expert parameters.
Company's own caveat: Reflection admits Kimi K3 still leads on raw capability, positioning Beam's advantage as inference-time efficiency.

Training ledger

Pretraining: 23.8T tokens from web and proprietary licensed data; 6,144 GB300 NVL72 GPUs, under 4 weeks. Derived estimate (not company-disclosed): ~4.13M GPU-hours (6144 × 28 × 24).
Data curation: ~95% of raw web tokens discarded after parsing/dedup; the pipeline recovered ~1.8T high-quality tokens, 87% of them hand-picked web code (~7.6% of total corpus).
Architecture: 52 layers, bounded residual-stream RMS throughout pretraining via depth-based scaling, SandwichNorm, per-element attention gating, and FP32 residual accumulation. Nine semi-automatic rollbacks for gradient-norm spikes and suspected silent data corruption; late-training goodput 92.3%.

| Stage | Scale | Duration | |---|---|---| | Pretraining | 6,144 × GB300 NVL72 | < 4 weeks | | High-compute RL | 10.5K GB300 | 4 weeks | | Sandbox environments | ~1M coding & STEM environments | 1.3B cumulative calls |

Combined estimate: ~11.2M GPU-hours, with RL accounting for roughly 60%. Beam is a text-only model; other modalities must be converted to text first.

Benchmarks: the price of efficiency

| Benchmark | Beam | Kimi K3 | Gap | |---|---|---|---| | DeepSWE v1.1 | 44.4 | 68.0 | 23.6 | | SWE Bench Pro v2-Hard | 77.2 | 88.2 | 11.0 | | Terminal Bench v2.1 | 80.1 | 88.3 | 8.2 | | SWE Atlas repo QA | 34.6 | 68.0 | 33.4 |

Beam trails on all four. The largest deficit is repository QA (33.4), roughly 4x the Terminal Bench gap—suggesting a context/retrieval training mix issue rather than a parameter-count issue.

On efficiency: vs GLM-5.2, Beam claims similar scores on advanced reasoning benchmarks with 3–4x less inference compute, computed as FLOPs ≈ 2 × active params × avg tokens generated per attempt. The model card explicitly excludes prefill, context-dependent attention, and serving overhead, and labels this an approximate comparison, not a measured inference cost.

Making "staleness" trainable

100M+ rollouts in high-compute RL: 80M on Beam's reasoning experts, 30M on Inkling, 753K on MiMo (~106 : 40 : 1).
Asynchronous policy gradients: long rollouts span multiple checkpoints, making early tokens stale relative to the current policy; numeric drift between training and inference engines compounds this. Fix: track the model version per token and compensate for policy staleness at the algorithm level. Reported stable even when interaction data lags by one day (107 weight versions).
Infrastructure: 110K concurrent rollouts on average; inference:training GPU ratio tuned between 3.9:1 and 5.4:1; median 12s for new weights to reach the inference cluster; 75% inter-rack traffic reduction; 2.2x faster team-wide weight adoption. 71 inference incidents handled without interrupting training; median recovery 8 minutes; lost capacity 0.02% of served GPU-minutes.

The sandbox factory

The least flashy and most telling section: up to 170K concurrent sandboxes, over 1 billion sandbox creation requests, across 20+ clusters, two clouds, four regions. Derived: ~1,000 reuses per environment on average. 90% of new sandboxes ready within 10 seconds; dynamic packing kept training batches 99.99% full; per-GPU throughput variance under 1.5%, sustained even as average rollout length grew ~70%. This scale of programmatic environment generation is what makes 100M rollouts feasible at all.

What the company itself concedes

Original quote: with frontier open models like Kimi K3 still ahead on raw capability, Beam's advantage is inference-time efficiency.
Half of the efficiency-comparison numbers come from public data by Artificial Analysis and DataCurve, not from the vendors themselves.
The 100M-rollout scale claim is hedged: "we believe this is among the largest by a public lab."
Safety evaluations were pending at publication.

Safety training uses a separate pipeline: starting from a pretraining checkpoint, independent SFT and RL, merging two teachers—one large-scale RL teacher, one dedicated to safety alignment—via multi-teacher on-policy distillation. Correlations: r = 0.79 for predicting RL gains, but only r = 0.46 for predicting Best-of-N ceilings.

Weights are planned under Apache 2.0 within the month, alongside a technical report, model card, and a full runtime evaluation/fine-tuning stack. On release day the model was still in final red-teaming, available only to a small waitlist. The accurate status of "open weights" at that moment: announcement issued, weights not yet delivered.

Sources and caveats

All architecture, training, and benchmark figures come from Reflection's October 5 official blog post (introducing-beam); no technical report was available for verification.
The 4.13M and 11.2M GPU-hour figures are derived from stated GPU counts and durations, not company-disclosed numbers.
The 3–4x inference-compute claim is the blog's own estimate; its formula and exclusions are transcribed as stated and self-labeled as approximate.
Beam is text-only; other modalities must be converted to text first. Weights, license terms, and the technical report were not public on the release day referenced.