Nvidia-backed Reflection AI has launched its first open-weight artificial intelligence model, Beam, as the US-based startup looks to compete with lower-cost Chinese AI models in coding and agentic tasks.
The model was unveiled on October 5, 2026. Reflection AI said Beam is competitive with Chinese AI startup Z.ai’s GLM-5.2 and is approaching Qwen3.8-Max in coding and agentic tasks. The launch comes as US technology companies face growing competition from Chinese open-weight models that offer lower costs, greater customisation and strong coding capabilities.
Beam contains 501 billion total parameters, which are variables that determine how an AI system processes information. However, the model activates only 23 billion parameters for each task. Reflection AI said this approach allows Beam to use only part of its network for individual tasks, helping make the model faster and cheaper to run.
By comparison, Z.ai’s GLM-5.2 has about 744 billion total parameters and 40 billion active parameters.
Reflection AI’s model is based on a sparse Mixture-of-Experts architecture and has been developed specifically for coding, reasoning and agentic workloads. The company said Beam is currently undergoing final red-teaming and evaluations, with an early version available to a select group of users.
Reflection AI said it plans to release Beam’s weights, technical report, model card and developer materials later in October. The company has also said the model weights will be released under an Apache 2.0 licence.
The company’s own evaluation shows Beam delivering performance comparable to GLM-5.2 on advanced reasoning benchmarks while using three to four times less inference compute. Reflection AI said the model is also designed to provide strong capabilities at lower cost, positioning it as a potential workhorse for enterprise coding and agentic workloads.
The model’s development involved large-scale reinforcement learning. Reflection AI said it used 10.5K NVIDIA GB300 GPUs for four weeks and generated more than 100 million rollouts during the training process. The company also said Beam was pretrained on 23.8 trillion tokens from the web and proprietary licensed datasets.
Reflection AI was founded in 2024 by former DeepMind researchers Misha Laskin and Ioannis Antonoglou. The company develops tools aimed at automating software development, an area where AI adoption has been growing rapidly.
Earlier this year, Reflection AI also signed a deal with SpaceX for additional computing capacity at the Elon Musk-led company’s Colossus 2 data centre.
The launch of Beam comes amid increasing competition between US and Chinese AI developers in the open-weight model segment. Chinese models such as DeepSeek, Kimi, GLM and Qwen have gained attention for their coding capabilities, lower operating costs and ability to be customised by users.
With Beam, Reflection AI is seeking to establish a stronger US presence in the open-weight AI model market while focusing on coding and agentic applications. The company’s comparisons with GLM-5.2 and Qwen3.8-Max place Beam directly against some of the leading Chinese open models in these areas.
