adtestbench

AI news · Reflection AI · models

Reflection unveils Beam, a 501B open-weight model

The Nvidia-backed lab says its first model nears GLM 5.2 on reasoning with three to four times less inference compute; weights follow later this month.

By Alexandre S. , 02:30 UTC

Reflection AI introduced Beam, its first model, on Monday 5 October 2026. It is a mixture-of-experts model with 501 billion parameters, 23 billion of them active per token, aimed at coding, reasoning and agent work. Beam is still in final red-teaming: Reflection says the weights, technical report and model card come later this month.

Retro-futurist illustration: a striped sunset over a grid horizon under a starry sky, with a satellite standing on the horizon.
Drawn by adtestbench from “Introducing Beam: Reflection's 501B open-weight model”,

The pitch is efficiency. Reflection says Beam scores close to Z.ai’s GLM 5.2 on advanced reasoning while using three to four times less inference compute, an estimate it calls approximate rather than a measured cost. Its own table has Beam slightly behind GLM 5.2 on most reasoning benchmarks and ahead on SWE-Bench Pro, 65.5 to 62.1, with Qwen 3.8-Max higher on both. Reflection also concedes that Kimi K3 leads on raw capability.

Reuters, in copy carried by Asharq Al-Awsat, frames Beam as a US answer to Chinese open models and notes the lab was founded in 2024 by former DeepMind researchers Misha Laskin and Ioannis Antonoglou. SiliconANGLE reports that Beam starts in an early access programme and that Reflection has raised money at a $25 billion valuation.

For teams that self-host, Beam adds a US-made open-weight option that sells on cost per answer rather than top scores, once the weights ship.