Breaking

Reflection AI Unveils Beam, a 501B Open-Weight Model for Coding and Agents

October 6, 2026 at 13:01 EDT

Foundation Models
Open Source
Software Dev & Coding

Reflection AI introduced Beam, its first open-weight model, positioning the 501-billion-parameter system as a Western entrant in a market lately led by Chinese open models. The company says Beam is built for coding, reasoning and agentic workloads, but the release remains a preview: public weights, a technical report, a model card and safety results are promised later in October rather than available today.

Oct. 5, 2026 · Reflection AI

Beam enters the open-weight race

Reflection AI unveiled a 501B-parameter sparse model for coding, reasoning and agentic workloads. Its pitch is more intelligence per active token in an open-weight package.

501B

total parameters

23B

active per token

23.8T

pretraining tokens

100M+

RL rollouts

Sparse scale, shown to proportion

Each block represents about 23B parameters. Beam carries 501B total parameters, while only 23B are active for each token.

501Btotal weights

23Bactive path

Takeaway: Beam's design separates total capacity from per-token compute.

Terminal Bench v2.1 snapshot

Published scores show Beam close to GLM 5.2 but behind Kimi K3 and DeepSeek V4.1 Flash on command-line work.

80.1Beam

88.3Kimi K3

90.6DeepSeek V4.1 Flash

Why builders care

Open weights, fine-tuning and local control could matter for coding agents, enterprise data and sovereign AI deployments.

What remains open

Independent benchmark results, public pricing and safety details were still unresolved at announcement.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.