# Reflection AI unveils Beam, a 501-billion-parameter open-weight model built to compete with China's GLM-5.2

The Nvidia-backed startup says Beam activates 23 billion parameters per token and scores 80.9 on SWE-bench Verified, a test of real software bug fixes. It also says Beam needs three to four times less inference compute than GLM-5.2, and it promises Apache 2.0 weights later in October.

- Published: 2026-10-06T06:03:25.241Z
- Canonical: https://polylog.news/ai/2026-10-06/reflection-ai-unveils-beam-a-501-billion-parameter-open-weig
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [Reflection AI (via Hacker News)](https://reflection.ai/blog/introducing-beam), [MarkTechPost](https://www.marktechpost.com/2026/10/05/reflection-ai-introduces-beam-a-501b-open-weight-moe-model-with-23b-active-parameters-for-coding-and-agentic-workloads/), [The Next Web](https://thenextweb.com/news/reflection-ai-beam-open-weight-model), [SiliconANGLE](https://siliconangle.com/2026/10/05/reflection-ai-debuts-open-source-beam-model-with-501b-parameters/), [IT之家](https://www.ithome.com/1/009/932.htm)

Reflection AI was founded in 2024 by Misha Laskin and Ioannis Antonoglou, both formerly of Google DeepMind. On October 5 it [introduced Beam](https://reflection.ai/blog/introducing-beam), its first open-weight model. Beam is a text-only sparse mixture-of-experts model with 501 billion total parameters, 23 billion of them active per token, and a 1-million-token context window. It was [pretrained on 23.8 trillion tokens](https://www.marktechpost.com/2026/10/05/reflection-ai-introduces-beam-a-501b-open-weight-moe-model-with-23b-active-parameters-for-coding-and-agentic-workloads/) of curated web data and proprietary licensed datasets. Access is limited to an early-access program for now. Reflection [says](https://siliconangle.com/2026/10/05/reflection-ai-debuts-open-source-beam-model-with-501b-parameters/) it will publish the weights under Apache 2.0 later this month, together with a technical report, a model card, and tools for evaluation and fine-tuning.

Reflection disclosed its compute budget. It pretrained the base model on 6,144 Nvidia GB300 GPUs in under four weeks. The reinforcement-learning stage then ran on [10,500 GB300s for four weeks](https://www.marktechpost.com/2026/10/05/reflection-ai-introduces-beam-a-501b-open-weight-moe-model-with-23b-active-parameters-for-coding-and-agentic-workloads/) and produced more than 100 million rollouts. Wccftech [reports](https://wccftech.com/nvidia-backed-reflection-used-10500x-gb300-gpus-and-46-4-million-sandboxes-per-day-for-4-weeks-to-train-beam-a-501b-open-weight-model-that-is-insanely-efficient/) that the RL environment ran 46.4 million sandboxes per day. The post-training budget was larger than the pretraining budget. That matches how labs now build agentic coding models, where most of the spending goes to RL against executable environments.

Reflection's own scorecard gives Beam [80.9 on SWE-bench Verified, 80.1 on Terminal-Bench 2.1 and 97.8 on the American Invitational Mathematics Examination (AIME) 2026](https://thenextweb.com/news/reflection-ai-beam-open-weight-model). Against Z.ai's GLM-5.2, Beam scores 65.5 to GLM-5.2's 62.1 on SWE-bench Pro v1, and 80.1 to GLM-5.2's 81.0 on Terminal-Bench 2.1. AIME is near its ceiling at this tier and tells readers little. On the agentic coding evaluations that separate models, the published numbers show a tie with GLM-5.2. They do not show a lead.

The efficiency claim is the basis of the launch, and it is the weakest part of the case. Reflection says Beam matches GLM-5.2's reasoning scores with three to four times less inference compute. The Next Web [notes](https://thenextweb.com/news/reflection-ai-beam-open-weight-model) that the figure is an estimate across DeepSWE, Humanity's Last Exam and Terminal-Bench 2.1, and that Reflection itself calls it an approximate comparison rather than a measured cost. Nobody has verified any of the scores independently. A 23-billion-active design does reduce per-token FLOPs. Serving still needs about 501 GB for the weights alone at 8-bit precision, so Beam remains a multi-GPU deployment and will not run on a single node.

The framing is geopolitical. Chinese-language technology media [reported the launch](https://www.ithome.com/1/009/932.htm) as a US challenge to DeepSeek and Kimi, and Taiwanese outlets [asked whether Beam can compete with GLM and Qwen](https://techorange.com/2026/10/06/reflection-ai-beam/). Reflection belongs to Nvidia's Nemotron Coalition of US open-model builders. It has signed a deal with South Korea's Shinsegae Group, a conglomerate, for a [250-megawatt data center](https://www.shopifreaks.com/reflection-seeks-2-5b-raise-at-25b-valuation-as-nvidia-backed-open-source-ai-startup-targets-sovereign-ai/) that will run Korean-language models. It is reported to be raising $2.5 billion at a $25 billion pre-money valuation, meaning before the new money is counted. That compares with the $8 billion valuation at which Nvidia invested about $800 million.

## What this means

Beam competes with Chinese labs on provenance, not on capability. Z.ai, Alibaba's Qwen team, DeepSeek and Moonshot have supplied the default downloadable model to governments and enterprises that cannot or will not depend on US closed APIs. Beam offers those buyers a roughly equivalent model with a US origin and a permissive license, at lower active compute. Nvidia benefits in two ways. It supplied the GB300 training fleet, and it is building an allied open-model ecosystem that keeps sovereign data center buyers on its hardware. Chinese labs lose export reach, mostly in markets aligned with the US, such as Korea. The precedent is DeepSeek-R1 in January 2025. It triggered a one-day sell-off in Nvidia shares, and Washington then explicitly promoted US open-weight models. Beam is a commercial product of that policy. The market has priced Reflection as a strategic asset, as the reported increase from $8 billion to $25 billion shows. What it has not yet shown is whether the model performs as well on evaluations that Reflection did not run. The likelier outcome is that independent SWE-bench Verified and Terminal-Bench runs fall within a few points of Reflection's numbers in the weeks after the weights ship, and that the three-to-four-times compute advantage shrinks once serving cost is measured. That would still make Beam the default US open-weight option for sovereign deployments within months. This depends on the weights shipping in October under Apache 2.0 as promised. A delay, a license narrower than Apache 2.0, or independent scores clearly below GLM-5.2 would prove the read wrong.

## What to watch

- Whether the weights and technical report ship in October under Apache 2.0, because the claim that Beam is a US alternative to Chinese open models depends on it actually being downloadable.
- Independent SWE-bench Pro and Terminal-Bench reproductions once weights are public, since the published edge over GLM-5.2 is within a few points and comes only from Reflection.
- Whether more Asian or Gulf conglomerates follow Shinsegae and sign data center deals tied to Reflection models, which would show that the US open-model offer is winning sovereign buyers away from Chinese stacks.