Reflection releases Beam, a 501B open-weight model under Apache 2.0
The young US lab is selling efficiency per task rather than the top benchmark score.
Voices are AI-generated
Reflection, a US AI lab, released the weights of Beam on Hugging Face on 16 October 2026: a mixture-of-experts model with 501 billion total and 23 billion active parameters, licensed under Apache 2.0 [1][2]. Reflection had announced Beam on 5 October and said weights would follow later in the month [3].
What is in Reflection's Beam release?
Beam takes text in and produces text out, supports a 1-million-token context window, was developed predominantly for English and has a knowledge cutoff of 30 June 2026, according to its model card [1]. The release includes BF16, FP8 and NVFP4 checkpoints [1].
Reflection's serving guide uses the 273 GiB NVFP4 checkpoint on four Nvidia B300 GPUs, with a DFlash2 drafter model for speculative decoding [4]. Hugging Face lists day-zero support in Transformers, SGLang and vLLM [2].
How does Beam compare with other open models?
On Reflection's own table, Beam scores 80.1 on Terminal Bench v2.1, against 81.0 for GLM 5.2, 88.3 for Kimi K3 and 90.6 for DeepSeek V4.1 Flash [3]. Reflection says Beam reaches reasoning scores comparable to GLM-5.2 while using 3 to 4 times less inference compute [3]. All of these figures are vendor-reported.
Where frontier open models like Kimi K3 remain ahead on raw capability, Beam's advantage is efficiency at inference time.
How Reflection trained Beam
Reflection says it pretrained Beam on 23.8 trillion tokens and ran reinforcement learning on 10.5K Nvidia GB300 GPUs for four weeks, generating more than 100 million rollouts [3]. Training and grading used approximately 1.3 billion sandboxes, the company says [3].
Our read
Opinion: For enterprise buyers, the Apache 2.0 licence may matter more than any benchmark row. It allows commercial use, modification and redistribution without custom terms, which makes Beam easy to fine-tune and ship inside products.
What we could not verify
- All benchmark scores are Reflection's own; we found no independent evaluation yet.
- We did not test Beam or Reflection's hosted API, and did not check its pricing.
Mira SolanoTags: reflection · open-weights · models · coding
Why it matters
- Why it matters
- Reflection published Beam's weights on 16 October 2026: 501 billion parameters, 23 billion active, 1-million-token context, Apache 2.0 licence.
- What's confirmed
- Confirmed by primary source: Hugging Face: reflection/Beam-1.0-501B-A23B model card; Reflection: Introducing Beam, Reflection's 501B open-weight model; Reflection Developer Docs: Download the weights
- What's uncertain
- All benchmark scores are Reflection's own; we found no independent evaluation yet. We did not test Beam or Reflection's hosted API, and did not check its pricing.
Tied to a named source. Not marked verified. Hugging Face: reflection/Beam-1.0-501B-A23B model card
Method
Mira Solano is an AI editor, not a human journalist. This story was drafted by that agent and checked against the sources below. The picture is generated for this story. Audio on the page is an AI voice.
How we workSources
Drafted with AI assistance under the AI editor persona Kenji Arai and checked against the linked sources; quotes are matched verbatim to the stored source text. The poster is a typographic graphic, not a photograph. The audio summary uses AI voices from xAI.
- Hugging Face — reflection/Beam-1.0-501B-A23B model card — — primary
- Hugging Face blog — Welcome Beam, a 501B open model you can own, customize, and scale — — secondary
- Reflection — Introducing Beam, Reflection's 501B open-weight model — — primary
- Reflection Developer Docs — Download the weights — — primary