Reflection AI, an American startup, has introduced Beam, its inaugural model featuring open weights, comprising 501 billion parameters and activating 23 billion per query.

Introducing Beam: a highly efficient agentic open model with 501B total parameters and 23B active.
— Frontier reasoning efficiency
— Advances the Western open frontier on coding & agentic tasks
— Trained end-to-end from scratch
Full weights release this month.
Learn more about Beam: https://t.co/c3Qx2cpM8G

— Reflection (@reflection_ai) October 5, 2026

The company is focusing on programming and AI agent functionalities, claiming that Beam is on par with the Chinese GLM-5.2 but with significantly lower inference costs.

U.S. Response to Chinese Open-Weight Models

Beam marks the first model from Reflection AI that extends beyond the company’s internal projects. The startup positions it as a competitor to Chinese open models, including Z.ai's GLM-5.2 and Alibaba's Qwen3.8-Max.

According to Reflection, Beam delivers comparable outcomes to GLM-5.2 and is closing the gap with Qwen3.8-Max in programming tasks and autonomous AI agent operations.

The model utilizes a Mixture of Experts (MoE) architecture. While Beam consists of 501 billion parameters, only about 23 billion are active during a single query, which helps reduce computational costs compared to similar models.

Reflection claims that in advanced reasoning tests, Beam performs at the level of GLM-5.2 while using 3 to 4 times fewer computational resources for inference. However, the company acknowledges that the Chinese Kimi K3 currently surpasses Beam in absolute capabilities.

Training Conducted on 10,500 Nvidia Accelerators

To train Beam, Reflection leveraged a large-scale Nvidia infrastructure. The high-load reinforcement training phase lasted four weeks and utilized 10,500 Nvidia GB300 GPUs, conducting over 100 million runs.

Additionally, around 1.3 billion virtual environments were employed for training and validating the model. The maximum context length during this phase reached 256,000 tokens, and the system could support up to 110,000 parallel runs simultaneously.

The base version of Beam was pre-trained for less than four weeks on a cluster of 6,144 Nvidia GB300 NVL72, utilizing 23.8 trillion tokens from web sources, public data, and licensed datasets.

Beam Currently Available via Waitlist

Beam is currently undergoing final security checks. An early version of the model is being offered to a limited group of pre-registered users.

In the coming weeks, Reflection plans to release the model weights, a technical report, a neural network passport, and developer tools. The weights will be distributed under the Apache 2.0 license, allowing users to run, fine-tune, and integrate the model into their own systems.

It is worth noting that in July, Nvidia, Meta, Microsoft, and 22 other American companies urged U.S. authorities not to impose restrictions on open-weight models.

Follow ForkLog on social media:

Telegram (main channel) Facebook X