Reflection Releases Open Model Beam, Targeting Chinese Rivals: what are the key facts?

Reflection AI released Beam, its first frontier open-weight model, claiming reasoning performance on par with leading Chinese open models.

Reflection Releases Open Model Beam, Targeting Chinese Rivals: what are the key facts?

Beam is a text-based mixture-of-experts model with 501 billion total parameters and 23 billion active parameters.

Reflection Releases Open Model Beam, Targeting Chinese Rivals: what are the key facts?

The model was pretrained on 23.8 trillion tokens and supports a context window of 1 million tokens.

Reflection Releases Open Model Beam, Targeting Chinese Rivals: what are the key facts?

Reflection claims Beam uses 3 to 4 times less inference compute than competitors, though the performance claims have not been independently verified.

Reflection Releases Open Model Beam, Targeting Chinese Rivals: what are the key facts?

The company plans to release the model weights and technical details this month and distribute them through hyperscale cloud providers.

Physix Frontier · News Briefing Card (enbrief · Oct 6, 2026)

Reflection Releases Open Model Beam, Targeting Chinese Rivals

KEY FACTS

Reflection AI released Beam, its first frontier open-weight model, claiming reasoning performance on par with leading Chinese open models.
Beam is a text-based mixture-of-experts model with 501 billion total parameters and 23 billion active parameters.
The model was pretrained on 23.8 trillion tokens and supports a context window of 1 million tokens.
Reflection claims Beam uses 3 to 4 times less inference compute than competitors, though the performance claims have not been independently verified.
The company plans to release the model weights and technical details this month and distribute them through hyperscale cloud providers.

KEY DATA

501 billionTotal parameters

23 billionActive parameters

23.8 trillionPretraining tokens

1 million tokensContext window

PHYSIX OBSERVATION

The competition in open models is shifting from parameter scale to inference cost. If Beam really can save several times the compute, the barrier to on-premises enterprise deployment will drop noticeably. But the performance claims have not been verified by a third party, and with Nvidia investment and compute ties behind it, this looks more like a positioning battle around the AI factory ecosystem. Users would do well to wait for the weights to be released before making a judgment.

Source: enbrief original report ↗