Reflection AI, a U.S. startup backed by Nvidia, said on Oct. 5 that it has launched Beam, its first open-weight model, positioning it for software development, reasoning, and AI agent workloads.
Weights set for release under Apache 2.0
According to Reflection, Beam uses a sparse mixture-of-experts, or MoE, architecture with 501 billion total parameters and 23 billion active parameters per inference pass. The company said the model is in final red-team testing and evaluation, and that it plans to release the weights later this month under an Apache 2.0 license together with a technical report and model card.
Benchmark claims against Chinese open models
Reflection said Beam posts advanced reasoning benchmark scores comparable to GLM-5.2, an open model from Z.ai, while requiring 3x to 4x less inference compute. It added that the efficiency gap is even wider when Beam is compared with models above 2 trillion parameters, including Alibaba’s Qwen 3.8-Max.
On coding and agent tasks, the company described Beam as competitive with larger open models such as GLM 5.2 and said it comes close to Qwen 3.8-Max. At the same time, Reflection acknowledged that frontier open models including Moonshot AI’s Kimi K3 still lead on raw capability, while Beam’s edge lies in inference efficiency.
Published scores and outside scrutiny
In the results released by Reflection, Beam scored 80.9 on SWE-bench Verified and 80.1 on Terminal Bench v2.1. In the same table, Kimi K3 posted 88.3 on Terminal Bench v2.1, while GLM 5.3 scored 88.2.
TechCrunch reported that those results have not been independently verified.
Training scale and reinforcement learning run
Reflection said Beam was pretrained on 23.8 trillion tokens. That pretraining run was completed in under four weeks on 6,144 Nvidia GB300 NVL72 GPUs. The reinforcement learning phase that followed used 10,500 GB300 GPUs over four weeks, generated more than 100 million rollouts, reached a maximum context length of 256,000 tokens, and used about 1.3 billion sandbox environments for training and evaluation.
The company described it as one of the largest reinforcement learning training runs carried out by an open lab to date, and said capability continued to improve as compute increased, with no sign of a plateau.
Context window and safety setup
Beam is a text-only model with a context window of up to 1 million tokens. Users can adjust a “reasoning intensity” setting to trade off between response length and performance.
On safety, Reflection said it trained a separate model focused on safety and alignment, then merged it with the large-scale reinforcement learning model. The company said safety evaluation results will be included in the technical report.
Funding, valuation, and target customers
TechCrunch said Reflection was founded in 2024 by two former Google DeepMind researchers. PitchBook data cited in the report shows the company has raised about $4.7 billion from investors including Nvidia, Sequoia Capital, and Lightspeed, with a most recent pre-money valuation of $25 billion.
TechCrunch also reported that Reflection signed compute contracts this summer worth more than $7 billion in total with SpaceX and Nebius, securing access to Nvidia GB300 chips through 2029.
According to TechCrunch, Reflection is targeting Beam and later models at enterprise and sovereign AI customers. The pitch is that institutions can train Reflection models on their own data and build localized “AI factories.” The company has already tested a sovereign AI factory partnership with South Korea’s Shinsegae Group. Reflection also said Beam is the first model in a broader series and that the next one is already in training.
