> ## Content Index > Fetch the complete content index at: https://globalfeed.ai/llms.txt > Use this file to discover other available public pages before exploring further. # Reflection AI introduces Beam, a 501B-parameter open-weight model - URL: https://globalfeed.ai/en/reflection-ai-introduces-beam-a-501b-parameter-open-weight-model/ - Published: 2026-10-07T15:01:12.000Z - Updated: 2026-10-07T15:01:12.000Z - Description: Reflection AI introduced Beam, its first open-weight model, with 501 billion total and 23 billion active parameters, trained for coding, reasoning and agentic work. According to the company, the weights will be released this month under Apache 2.0; for now, early access runs through a waitlist. - Author: GlobalFeed Editor - Tags: Reflection AI, Beam, open models, coding, ai agents, x-reflection_ai, dil-en, elle **Reflection AI** introduced **Beam**, its first open-weight model, on October 5, 2026\. According to Reflection's announcement, Beam is a sparse Mixture-of-Experts (MoE) model with 501 billion total parameters, of which 23 billion are active at each step. The company says it trained the model with a particular focus on coding, reasoning and agentic workloads, where AI agents carry out multi-step tasks on their own. Beam is currently going through final red-teaming and evaluations, and early access is opening to a select group of users who sign up for a waitlist. According to Reflection, the weights will be released this month under the Apache 2.0 license, together with a technical report and a model card, so organizations will be able to run Beam on their own servers. As of the announcement, the weights are not yet available for download. ## Beam works with text and supports long context Beam is a text-only model; according to the company, it can work with information from other modalities, such as images, only when that information is represented as text. Reflection says its midtraining stage extends the model's effective context length, the amount of text it can take into account at once, to 1 million tokens. Users can set a reasoning effort parameter: lower settings favor shorter responses, while higher settings allow longer reasoning on demanding tasks. ## Training relied on large-scale reinforcement learning According to the announcement, Beam was pretrained on 23.8 trillion tokens drawn from the web, public sources and proprietary licensed datasets. Pretraining finished in under four weeks on a cluster of 6,144 NVIDIA GB300 NVL72 GPUs. In the reinforcement learning (RL) stage that followed, 10,500 NVIDIA GB300 GPUs ran for four weeks and generated more than 100 million rollouts. Reflection says it built nearly one million task environments for this stage, covering areas such as software engineering, terminal use, web search and STEM. The company says capabilities kept improving until the end of the run, with no sign of a plateau. ## Benchmarks published by Reflection Reflection shared tables comparing Beam with GLM 5.2, GLM 5.3, Kimi K3, Qwen 3.8 Max, DeepSeek V4.1 Flash, Nemotron 3 Ultra and Inkling. In the company's tables, Beam scores 80.9 on SWE-bench Verified, 80.1 on Terminal Bench 2.1 and 90.5 on GPQA Diamond. The same tables list GLM 5.2 at 81.0 and Kimi K3 at 88.3 on Terminal Bench 2.1; on SWE Bench Pro v1, Beam's 65.5 sits between GLM 5.2 at 62.1 and Qwen 3.8 Max at 67.7\. Reflection describes Beam as competitive with GLM 5.2 and approaching Qwen 3.8 Max on coding and agentic tasks; it notes that models such as Kimi K3 remain ahead on raw capability and places Beam's advantage in efficiency at inference time. According to the company, Beam reaches scores comparable to GLM 5.2 on advanced reasoning benchmarks while using 3 to 4 times less inference compute. Reflection notes that this comparison is an approximate estimate based on active parameter count and generated tokens, and that it excludes prompt processing and serving overhead. ## Weights and release plan Reflection says it will release the weights along with documentation and the full stack for running, evaluating and fine-tuning the model. The company plans to launch Beam with a group of distribution partners and integrations with a broad range of open-source libraries; the partners are not named in the announcement. Sign-ups for early access are taken on Reflection's platform page. The results of the safety evaluations will appear in the technical report, and the company says it will also open-source the safety evaluations it developed and used internally. Reflection describes Beam as the first model in a series and says it is already training the next one. ## About Reflection AI According to its website, Reflection AI aims to make intelligence open and accessible to all. The company says it builds AI systems that give enterprises, governments and sovereign entities control over their intelligence, and it makes three commitments: releasing open weights for its models, publishing the research behind them in technical reports, and open-sourcing software, including reinforcement learning tools and environments. According to the site, Reflection is validated on Dell AI Factory with NVIDIA and is building a sovereign AI cloud in South Korea with Shinsegae for enterprises and public agencies. All the published comparisons were prepared by Reflection; according to the company, scores for the other models come from Artificial Analysis and DataCurve data, and many results are missing for some models in the tables. Since the weights have not been released yet, Beam cannot be tested independently for now. The announcement does not state the hardware needed to run the model; running a 501-billion-parameter model on your own servers takes multiple data center GPUs. The announcement also does not say which human languages Beam supports or how well, so it is worth testing it on your own examples before using it on non-English text. For a fuller picture, follow evaluations from independent researchers as well. Sources: [Reflection AI: Introducing Beam](https://reflection.ai/blog/introducing-beam?ref=globalfeed.ai), [Reflection AI: About](https://reflection.ai/about?ref=globalfeed.ai), [Reflection Platform: early access](https://platform.reflection.ai/?ref=globalfeed.ai).