Signals›Open-source AI
Open-source AIShipping

Reflection's Beam is out under Apache 2.0 — a US open model aimed squarely at the Chinese leaders.

A genuinely permissive license on a 501-billion-parameter model is the real news. The benchmark claims still need someone else to check them.

ZN ZilverNet EditorialPublished · 5 min read · 4 sources Weights released Oct 16, 2026
The Beam wordmark beside concentric dotted circles on dark green
Image credit: Reflection
The short versionShipping
What it is
A text-only mixture-of-experts model: 501B parameters, 23B active, a 1M-token context, pretrained on 23.8 trillion tokens.
How far along
ResearchPrototypePilotShippingEverywhere
Shipping. Weights on Hugging Face since Oct 16 in BF16, FP8 and NVFP4.
The license
Apache 2.0 — no revenue threshold, no separate agreement. The model card says it is "intended for commercial and research use."
Still unproven
Reflection's performance claims "haven't been independently verified," as TechCrunch put it.

Announced, then shipped

Reflection announced Beam on October 5 and promised the weights within the month. On October 16 it delivered: "Today, we have released Beam's model weights on Hugging Face under an Apache 2.0 license." The company calls it "our first open-weight model."

Beam is available in three precisions plus a small "drafter" model that speeds up generation, through Reflection's own API, partner platforms, and on-premises through Dell and HPE.

Why the license is the story

Apache 2.0 is one of the most widely used permissive licenses in software. In practice it means you can use, modify and sell what you build without asking anyone's permission or crossing a revenue line.

That is a meaningful contrast with Meta's Muse Model License, released the day before Beam's weights, which requires a separate agreement for large model-service providers and an attribution label.

The claims, labeled as claims

  • Scores on par with Z.ai's GLM-5.2 on advanced reasoning benchmarks.
  • Outperforms leading Western open models while using "3-4x less inference compute."
  • Beats Inkling on four coding tests — though, as TechCrunch notes, Inkling handles images and Beam handles only text.

All three come from Reflection's own testing.

Who's behind it

Reflection was founded in 2024 by two former Google DeepMind researchers and is based in Brooklyn. It has raised about $4.7 billion according to PitchBook, with its last round at a $25 billion pre-money valuation; backers include Nvidia, Sequoia and Lightspeed.

It is spending at scale. Reflection says Beam was pretrained on 6,144 GB300 NVL72 GPUs in under four weeks, and it has compute deals worth more than $7 billion with SpaceX and Nebius running through 2029. It is also piloting a sovereign "AI factory" with Shinsegae Group in South Korea.

Update log
  • Published.