Translating into Español. The English version is shown until it is ready.
Reflection AI unveils Beam, a 501B open-weight model for coding and agents
En breve
US startup Reflection AI announced Beam, an open-weight Mixture-of-Experts model with 501 billion total parameters and 23 billion active parameters, built for coding, reasoning and agent workloads. The model is still in final red-teaming and evaluation, with early-access registration open; the company says weights will be published in October under the Apache 2.0 licence. Reflection says Beam matches larger open models such as GLM 5.2 at one-third to one-quarter of the inference compute, an estimate that has not been independently verified.
Leer la noticia completa 3 min de lectura
Reflection AI, a US startup based in Brooklyn, New York, announced an open-weight AI model called Beam on October 5 US time, according to Mynavi News, Impress Watch and GIGAZINE. The model is a Mixture-of-Experts design with 501 billion total parameters and 23 billion active parameters, and the company says it is built for coding, reasoning and agent-based workloads. The documents say Beam is in its final red-teaming and evaluation stage and that early-access registration is open. Its weights are to be published in October under the Apache 2.0 licence, together with a technical report, a model card and software for running, evaluating and fine-tuning the model, Mynavi News reported. [ 1 , 2 , 3 ]
The company published comparison figures for Beam. Mynavi News reported scores of 80.9 on SWE-bench Verified, 80.1 on Terminal Bench v2.1 and 44.4 on DeepSWE v1.1. On Terminal Bench v2.1 that is roughly level with GLM 5.2 at 81.0 but below Kimi K3 at 88.3 and DeepSeek V4.1 Flash at 90.6, while on DeepSWE v1.1 Beam is close to GLM 5.2's 44.0 but behind DeepSeek V4.1 Flash's 74.2. GIGAZINE described Beam as reaching parity with larger open models such as GLM 5.2 and approaching Qwen 3.8-Max on coding and agent tasks, while still behind frontier open models like Kimi K3 in raw capability. [ 1 , 3 ]
Reflection's central claim is efficiency at inference time. The company says Beam matches GLM 5.2 on advanced reasoning benchmarks while using about one-third to one-quarter of the inference compute, and that it competes with GLM 5.2 on some coding and agent benchmarks while approaching the larger Qwen 3.8-Max. Mynavi News reported that this figure is not a measurement: it is an estimate derived from Artificial Analysis and DataCurve data using active parameter counts and average generated token counts, and it excludes prompt preprocessing, context-dependent attention and model-serving overhead, so it is not a direct comparison of real inference costs. Mynavi News also said the performance figures published by Reflection have not been independently verified by a third party. [ 1 , 3 ]
On training, Mynavi News reported that Beam is a text-only model pretrained on 23.8 trillion tokens selected from the web, public data and licensed proprietary datasets, with context extended to up to 1 million tokens during mid-training so it can handle long code repositories and documents. The company said reinforcement learning used 10,500 NVIDIA GB300 units over four weeks, generating more than 100 million rollouts, with about 1.3 billion sandboxes and roughly 1 million training environments for coding, agent and STEM work. GIGAZINE reported the same GPU and rollout figures. Reflection said it saw no sign of performance plateauing as RL compute increased within the range it evaluated, and introduced a “reasoning effort” parameter so users can trade token use against harder-problem performance. [ 1 , 3 ]
For safety and alignment, the company said it trained separate models with supervised fine-tuning and reinforcement learning pipelines from pretrained checkpoints, then merged their capabilities into one model using “multi-teacher on-policy distillation”. Safety evaluation results are to be published in the technical report, and the company plans to open-source its internal safety evaluations. Mynavi News added that Beam is the first model in a series and that Reflection has already begun training a successor, with distribution partners and open-source library integrations to be available at launch. [ 1 ]
Mynavi News reported that Reflection AI was founded in 2024 by two researchers from Google DeepMind and is based in Brooklyn, New York, and has raised about $4.7 billion from investors including NVIDIA, Sequoia Capital and Lightspeed Venture Partners. The same outlet said the company signed contracts worth more than $7 billion in total with SpaceX and Nebius in the summer, securing GB300 capacity through 2029. These company-background details appear only in Mynavi News. [ 1 ]
Por qué importa
Beam is presented by its maker as a US open-weight answer to Chinese open models, and its claimed edge is inference efficiency rather than raw scores. If the efficiency holds up outside the company's own estimates, it could lower the cost of running coding and agent workloads — but the documents note no third-party validation of the performance or cost claims yet.
Datos clave
- Reflection AI announced the open-weight model Beam on October 5 US time. [ 1 , 2 , 3 ]
- Beam is a Mixture-of-Experts model with 501 billion total parameters and 23 billion active parameters. [ 1 , 2 , 3 ]
- The company says it is designed for coding, reasoning and agent-based workloads. [ 1 , 2 , 3 ]
- Reflection AI says the model is in its final red-teaming and evaluation stage, with early-access registration open. [ 1 , 2 , 3 ]
- Weights are to be published in October under the Apache 2.0 licence, alongside a technical report, model card and software. [ 1 , 2 , 3 ]
- Reflection reported pretraining on 23.8 trillion tokens and context extended to up to 1 million tokens. [ 1 , 3 ]
- Reflection reported benchmark scores of 80.9 on SWE-bench Verified, 80.1 on Terminal Bench v2.1 and 44.4 on DeepSWE v1.1. [ 1 ]
- Reflection says reinforcement learning used 10,500 NVIDIA GB300 units for four weeks and generated more than 100 million rollouts. [ 1 , 3 ]
Confirmado por varias fuentes
- Reflection AI announced an open-weight AI model called Beam. [ 1 , 2 , 3 ]
- Beam is a Mixture-of-Experts model with 501 billion total parameters and 23 billion active parameters. [ 1 , 2 , 3 ]
- The company says the model targets coding, reasoning and agent workloads. [ 1 , 2 , 3 ]
- The model is in final red-teaming and evaluation and early access is open for registration. [ 1 , 2 , 3 ]
- Model weights, a technical report and a model card are to be published in October. [ 1 , 2 , 3 ]
Aún sin aclarar
- Whether Beam's reported benchmark performance and efficiency hold up under independent testing. Mynavi News states that the performance figures published by Reflection have not been independently verified by a third party at this point.
- The exact date in October on which the weights and documentation will be released. The documents only say publication is planned for some time in October.
- How much inference compute Beam actually uses compared with GLM 5.2. Mynavi News reports the one-third to one-quarter figure is an estimate derived from Artificial Analysis and DataCurve data, not a measurement, and excludes prompt preprocessing, context-dependent attention and serving overhead.
- Reflection AI's funding and its contracts with SpaceX and Nebius. These details appear only in one document, Mynavi News.