The Information Machine
ConcludedRan 16 daysin the edition since 6 Oct 20266 sourcesupdated 21 Oct 2026

Reflection AI's Beam open-weight model

The gist

Latest: Reflection AI Releases Beam, a 501B Open-Weight MoE Model

Beam is the largest permissively licensed open-weight model released by a US lab to date, per Interconnects.ai. Benchmark comparisons show it trails several leading Chinese open-weight models on some tasks, sharpening ongoing questions about the competitive position of US open-weight efforts.

When stories like this move, I write about them in an email each morning.

One free email each morning, around 8 AM US Eastern.

The full picture

Reflection AI released Beam on October 16 under Apache 2.0, with open weights in BF16, FP8, and NVFP4 formats, a DFlash2 drafter, and the Mirror CLI training harness built with over 40 partners. Beam is a sparse mixture-of-experts model with 501B total and 23B active parameters per token. Pretraining used 23.8 trillion curated tokens on 6,144 GB300 NVL72 GPUs in under four weeks; the RL run used roughly 10,500 of the same GPUs for four additional weeks, producing over 100 million rollouts across nearly one million environments and about 1.3 billion sandboxes with no sign of a plateau. Reflection claims Beam achieves scores comparable to GLM-5.2 on advanced reasoning benchmarks while using three to four times less inference compute, though that comparison rests on estimated forward-pass FLOPs rather than measured end-to-end serving cost. Reflection targets enterprises and governments seeking to self-host, fine-tune with proprietary data, and avoid dependence on closed APIs. Reflection AI is backed by Nvidia.

How it developed
20 October 2026in the 21 Oct edition

Interconnects.ai published an assessment on October 20 of Beam, the 501-billion-parameter sparse MoE model Reflection AI released October 16, describing it as the strongest open model built in the US under a permissive license and second only to GPT-6-Astra in token efficiency.

The piece characterizes Beam's post-training stack as strong but does not call it the overall smartest open model available.

First citedInterconnects
16 October 2026in the 17 Oct edition

Reflection AI released Beam's weights on October 16 under the Apache 2.0 license, making the 501-billion-parameter sparse mixture-of-experts model available for self-hosting and fine-tuning.

The release includes BF16, FP8, and NVFP4 formats, a DFlash2 drafter, and Mirror CLI, the training harness used to build Beam, in collaboration with over 40 partners.

First citedDatabricks on X
14 October 2026

Anthropic's Guillaume Princen argued at the International AI Summit in Dublin that open-weight models carry more safety risk by definition

12 October 2026in the 13 Oct edition

October 12 reporting added a benchmark comparison not included in Reflection AI's headline chart: Beam scores 44.4 on DeepSWE 1.1 against the newer GLM-5.3's 61.0, extending the gap beyond earlier reported measures.

The reporting also clarified that Beam's weights and technical report, from the October 5 launch, remain planned for later in October and have not yet been published.

First citedImport AI
10 October 2026in the 11 Oct edition

The Financial Times reported October 10 that Nvidia is in early discussions to acquire Reflection AI or expand its existing $800M stake, with an acqui-hire structure, in which Nvidia would hire Reflection's staff and license its technology without triggering a full antitrust review, as one option under consideration.

The FT said a deal could materialize within weeks and described the Trump administration as supportive of Reflection AI as a competitor to Chinese AI efforts.

9 October 2026in the 10 Oct edition

Subsequent coverage on October 9 confirmed that Reflection AI is backed by Nvidia, a detail absent from the initial release reports.

Beam, released October 5, uses a sparse mixture-of-experts architecture with 501 billion total parameters and 23 billion active per token, trained on 23.8 trillion tokens and released under Apache 2.0. Reflection claims a three-to-four-times inference compute efficiency advantage over GLM-5.2, though one source notes the comparison uses estimated FLOPs rather than measured serving cost.

First citedTransformer News
7 October 2026in the 8 Oct edition

Coverage published October 7 framed Beam's market case around compliance and control rather than benchmark performance, citing Reflection CEO Misha Laskin's stated focus on regulated industries and governments that cannot use Chinese AI models.

A source also reported that Beam trails top Chinese open-source models and leading closed systems from OpenAI and Anthropic on benchmarks, placing the model's competitive argument around affordability and control rather than capability.

First in the edition · 6 October 2026
5 October 2026in the 6 Oct edition

Reflection AI released Beam on October 5, a 501-billion-parameter sparse mixture-of-experts model with 23 billion active per token under Apache 2.0, pretrained on 23.8 trillion tokens using 6,144 Nvidia GB300 GPUs.

A reinforcement learning phase on 10,500 GPUs generated over 100 million rollouts with no sign of a plateau. Reflection claims three-to-four times inference compute efficiency over GLM-5.2, though one source notes the figure rests on estimated forward-pass FLOPs rather than measured serving cost. On Terminal Bench v2.1, Beam scores 80.1 against GLM-5.2's 81.0, but DeepSeek V4.1 Flash at 90.6 and Kimi K3 at 88.3 do not appear in Reflection's headline benchmark chart.

20 July 2026

Analysis published showing GLM-5.2 and Kimi K3 achieve near-frontier performance at 17-35% of leading proprietary model costs

Sources
Milk Road AI on X
1 more source

Get the daily email

When stories like this move, I write about them in an email each morning.

One free email each morning, around 8 AM US Eastern.

See the latest edition · Privacy