Reflection releases Beam, a 501B open model under Apache 2.0

Major

Lab claim

On October 16, Reflection published Beam's weights. The model has 501 billion parameters, about 23 billion active, and a 1 million token context. The license is Apache 2.0.

How to read this claim

Measurement — No independent rerun of SWE-bench Verified, Terminal-Bench, or GPQA is cited.
Bill — List input and output per million tokens on the vendor page — a loop pays both again.
Door — No independent rerun of SWE-bench Verified, Terminal-Bench, or GPQA is cited.

Keep the habit · field card with prices and worked checks

Two open agent models

Figures from each lab's release page, October 15 and 16, 2026.
	Beam	Muse Spark Open
Lab	Reflection	Meta
Date	October 16	October 15
License	Apache 2.0	Muse Model License
Total parameters	501 billion	816 billion
Active parameters	23 billion	42 billion
Context	1 million tokens	1 million tokens
Runs as described	Hosts listed by Reflection	Two M5 Ultra Mac Studios, or a host

What changed

The day before, the open agent model a team could download was Muse Spark Open, under Meta's license, with 42 billion active parameters. Beam is the same kind of file under Apache 2.0, with 23 billion active.

What it introduces

A team can download an agent-class model under a license many companies already know how to approve. The coding scores that come with it are still Reflection's.

Who benefits Teams that want an open coding or agent model they can fine-tune. Anyone comparing Apache weights with Meta's Muse license.

Try this

If a coding agent is on the roadmap, run Beam against the repositories you already score, and read the Apache terms before you assume the weights can sit inside a product.

What we do not know

No independent rerun of SWE-bench Verified, Terminal-Bench, or GPQA is cited.

Evidence

Primary Reflection. Reported by Reflection. The benchmark table is Reflection's, run at temperature 0.7.

Independent No independent rerun is cited. The SWE-bench Verified, Terminal-Bench, and GPQA figures are Reflection's, measured at temperature 0.7.

Sources

Reflection Primary · October 16, 2026

The record

What happened

On October 16, 2026, Reflection released Beam, its first open-weight model, for coding, reasoning, and long agent jobs. The weights are on Hugging Face under Apache 2.0. Beam is a mixture-of-experts model with 501 billion parameters and 23 billion active, text in and text out, and a 1 million token context. Checkpoints ship in BF16, FP8, and NVFP4. Reflection's model card, evaluated at temperature 0.7, reports 81.6 percent on SWE-bench Verified, 80.1 percent on Terminal-Bench v2.1, and 90.6 percent on GPQA Diamond for the BF16 checkpoint. The FP8 and NVFP4 rows are close to those numbers. Hosts listed at launch include Amazon, Google Cloud, Microsoft Foundry, Baseten, Fireworks, Together, Modal, OpenRouter, and others.

Why it matters

Muse Spark Open, released the day before, is the open model with a custom license and a desktop story built around two Mac Studios. Beam is the open model with a license many companies already know how to approve, and a smaller active size, 23 billion against Spark Open's 42 billion. A team choosing an open coding model this week is choosing between those two, not between a blog and a rumor. Reflection's 81.6 percent on SWE-bench Verified is the number to put next to a test of your own repositories. It is Reflection's number. The card does not claim an independent rerun.

Reflection had described Beam earlier in the month and said the weights would follow. October 16 is the day the weights, the model card, and the safety report were posted.

The FP8 checkpoint is about half the disk of BF16: Reflection measures 469 GiB against 934 GiB. That is a serving detail, not a second model.