Reflection Beam: 501B Open-Weight MoE, Weights Due Soon

Introduction

On October 5, 2026, Reflection announced Beam, a 501B-parameter open-weight model, in a post titled "Introducing Beam: Reflection's 501B open-weight model." Beam is a sparse mixture-of-experts (MoE) text model with 23B active parameters, built for coding, reasoning, and agentic workloads. One point matters before anything else: the weights are not out yet. Reflection says "Beam is undergoing final red-teaming and evaluations," offers early access through a waitlist at platform.reflection.ai, and says it will release the weights, technical report, model card, and developer artifacts "later this month" under Apache 2.0. This article therefore treats Beam as an announced preview, and every performance figure below is Reflection's own.

Architecture and training

Sparse MoE means each token activates only a small share of the model's experts. With 23B of 501B parameters active, per-token compute is closer to a mid-sized dense model than to a 500B one, while total capacity stays large.

Reinforcement learning at scale

Reflection describes a high-compute RL phase: 10.5K NVIDIA GB300 GPUs for four weeks, over 100 million rollouts, a maximum context of 256K, roughly 1.3 billion sandboxes for training and grading, and nearly one million environments, with an average of 110K concurrent rollouts. Reflection says "we believe this is one of the largest scale RL runs conducted by any open lab to date," which is a statement of belief and not an independently established ranking.

Its fully asynchronous RL stayed stable even with one-day staleness, meaning rollouts were generated by a policy 107 weight versions behind the current one. A reasoning-effort parameter lets users trade response length against performance: lower settings favor shorter answers, and higher settings allow longer reasoning on demanding tasks.

Benchmarks: Reflection's table

Reflection's coding and agentic table uses NR for scores a model did not report. The figures below are Reflection's, and TechCrunch notes the company's performance claims "haven't been independently verified."

Reading the table honestly gives a mixed picture. Beam is ahead of Inkling and Nemotron 3 Ultra wherever they report, level with GLM 5.2 on DeepSWE (44.4 versus 44.0) and ahead on SWE Bench Pro v1, but slightly behind it on Terminal Bench (80.1 versus 81.0). It trails Qwen 3.8 Max on all three shared tests, and on DeepSWE and Terminal Bench it trails several listed models by wide margins. Reflection itself says Beam is "competitive" with GLM 5.2, "approaching" Qwen 3.8-Max on coding and agentic tasks, and that Kimi K3 "remains ahead on raw capability."

The efficiency claim

Reflection's main pitch is efficiency: on advanced reasoning benchmarks it says Beam scores comparably to GLM-5.2 while using 3-4x less inference compute. Its method estimates generation compute as 2 x active parameters x mean generated tokens, and Reflection states this excludes prompt prefill, context-dependent attention, and serving overhead, so it is "an approximate compute comparison rather than measured inference cost." TechCrunch reports GLM-5.2 has about 744B total and 40B active parameters, so Beam's smaller active count is part of the arithmetic.

Safety approach

Reflection says it trained a second model for safety and alignment and merged it with the main RL model through multi-teacher on-policy distillation, using deliberative alignment. Safety evaluation results are promised in the technical report, which is not yet published, so there is nothing to check yet.

Company context

TechCrunch describes Reflection as Brooklyn-based, founded in 2024 by two former Google DeepMind researchers, and reports it has raised roughly $4.7B per PitchBook from backers including Nvidia, Sequoia, and Lightspeed, with a last round at a $25B pre-money valuation. Compute deals with SpaceX and Nebius for GB300 access through 2029 reportedly total over $7B. Reflection is aiming at enterprises and sovereign "AI factories," and TechCrunch says it is testing a sovereign AI factory partnership with South Korea's Shinsegae Group. Inkling from Thinking Machines Lab is multimodal while Beam is text-only, and Beam outscores Inkling on the four coding tests where both report.

Outlook and assessment

The real test comes when the weights, model card, and technical report ship. Independent runs of the coding and agentic benchmarks, plus a measured cost comparison instead of a FLOPs estimate, will show whether the efficiency story holds in deployment. Teams that need an Apache 2.0 licensed model with long context and agentic tuning should watch for the release Reflection has promised for this month, but should not plan around Beam until the artifacts exist and third parties have tested them.

Editor's Verdict

Reflection Beam: 501B Open-Weight MoE, Weights Due Soon is a workable proposition that fills a clear gap, even if it doesn't fundamentally change the landscape.

The strongest case for paying attention: the planned Apache 2.0 license, if delivered as stated, would permit broad commercial use and modification of the weights. That alone raises the bar for what readers should expect in this space. Reinforcing that, a sparse design with 23B active parameters keeps per-token compute lower than the 501B total size suggests — practical value rather than just headline appeal. The broader signal worth registering is straightforward: the announcement is a preview, since Reflection says Beam is in final red-teaming and the weights, model card, and technical report arrive later this month. On the other side of the ledger, one constraint is real rather than a marketing footnote: the weights, technical report, and model card are not yet released, so none of the benchmark or safety claims can be independently reproduced today. It should factor into any serious decision. Layered on top of that, the coding table shows Beam trailing several listed models, including Qwen 3.8 Max, Kimi K3, and DeepSeek V4.1 Flash on DeepSWE and Terminal Bench — which narrows the set of teams for whom this is an obvious yes.

For multi-model deployment teams, cost-conscious operators, and developers willing to evaluate beyond the major labs, the smart move is to track its trajectory and revisit once the rough edges are filed down. For everyone else, the safer posture is to monitor coverage and revisit once the use cases that matter to your team are demonstrated in the wild.