Reflection AI Unveils Beam, an Open-Weight Model It Says Matches China's GLM-5.2

“DeepSeek and Qwen and all these models are our wake-up call because if we don't do anything about it, then effectively, the global standard of intelligence will be built by someone else,” Misha Laskin said in October 2025. “It won't be built by America.” Reflection AI's CEO had led reward modeling for Google's Gemini project at DeepMind before co-founding the company.

A year later, Reflection has a model, and only a select group of users can try an early version through a waitlist. Beam is in final red-teaming. Its weights are due “later this month” under Apache 2.0. Nobody can download them yet.

Announced Monday, October 5, Beam is a text-only model with 501 billion total parameters and 23 billion active per token. Reflection says its reasoning scores are comparable to Z.ai's GLM-5.2, released in June, at three to four times less inference compute, an estimate covering DeepSWE, Humanity's Last Exam and Terminal Bench 2.1.

The compute estimate

Parameters are the learned values that determine a model's responses. Beam activates only a portion for each token, a piece of text it processes or generates. In the October 2026 comparison, GLM-5.2 has roughly 744 billion total parameters and 40 billion active.

Open-weight means users can obtain those learned values and run the model on their own hardware. Laskin compared closed models to renting an apartment on the Sources podcast. “The only way to own intelligence is, by definition, if it’s open,” he said.

Reflection's efficiency calculation draws on Artificial Analysis and DataCurve data. It estimates generation compute as roughly twice the active parameter count multiplied by mean generated tokens per attempt, including reasoning and the final answer. Fewer active parameters and fewer generated tokens reduce that estimate.

The calculation excludes the work of processing the initial prompt, attention operations that vary with context, and serving overhead. Reflection calls it “an approximate compute comparison rather than measured inference cost.”

The benchmark gap

Beam is behind the newest Chinese models on most rows of Reflection's own table. All scores below are as reported by Reflection on October 5, 2026.

Reflection describes Beam as “competitive with” GLM-5.2 and “approaching” Qwen 3.8-Max. It acknowledges that models such as Kimi K3 “remain ahead on raw capability.”

Against Inkling, the Western open model from Thinking Machines Lab, Beam scores higher on the four coding tests where both report results. The systems have different scope: Beam is text-only, while Inkling is multimodal.

Artificial Analysis, an independent benchmarker given access by Reflection, is testing Beam. In an October 5 post, it said: “Early indicators suggest Beam will be one of the most token-efficient open models we've seen for its level of intelligence.”

That is an early indication from vendor-provided access. No independent evaluation of Beam has been published. No outside skeptic of Beam is named in the reporting.

How Beam was trained

Reflection says Beam was pretrained on 23.8 trillion tokens in under four weeks using 6,144 Nvidia GB300 GPUs. Reinforcement learning, where the model practices tasks and receives rewards for solving them, produced over 100 million rollouts, or task attempts, on 10,500 GB300 GPUs over four weeks. The company says it sourced nearly one million coding, agentic and STEM environments and believes this was among the largest reinforcement-learning runs by an open lab to date. Reflection says that during one reinforcement-learning phase on reasoning, software engineering and terminal tasks, Beam improved at web browsing even though the reinforcement-learning mix held no browsing tasks.

Can you run it at home?

Beam needs memory for all 501 billion parameters, though only 23 billion activate per token. Its weights remain unreleased, with no home-hardware guidance or compressed builds. A search of Hugging Face on October 5 found no Beam repository and no compressed version. The models with more than 400 billion parameters in those Mac Studio tests ran at 4-bit or lower, with 4-bit meaning each weight is stored in four bits.

In July 2026 Mac Studio measurements, an M3 Ultra with 512 GB unified memory ran 4-bit GLM-5.2 at 17.9 tokens per second, using 421.4 GB at peak. GLM-5.2 has 744 billion parameters, 40 billion active. Beam's smaller counts suggest it would likely fit in 512 GB at 4-bit and might run as fast or faster, if the software supports it. That is an inference, not a Beam measurement. Longer prompts slow generation and delay the first token; GLM-5.2 ran on experimental, unreleased software with roughly 3,000 prompt tokens and was not measured at long prompts.

Nvidia's 128 GB DGX Spark supports up to 200 billion parameters, or 405 billion with two linked. Beam exceeds those limits. The weights are due this month; whether a 4-bit build arrives and how it runs will be the next things to watch.

Buyers awaiting the weights

Laskin sees a market among enterprises and governments that cannot or will not use Chinese models but want to control their own AI systems. Those buyers “don't really have very good options today,” he said.

Reflection pitches what it calls “AI factories” to these institutions: customized local systems built by training its models on their proprietary data. The company plans distribution through cloud providers and integrations with open-source libraries.

Several large Western companies have recently embraced Chinese models to cut costs on repetitive tasks such as customer service and writing routine code. Airbnb CEO Brian Chesky said in 2025 that his company was “relying a lot” on Alibaba's Qwen, calling it “very good” and “fast and cheap.”

Clem Delangue, co-founder and CEO of Hugging Face, the platform for AI builders, said at the time of the October 2025 round: “This is indeed great news for American open-source AI.” He then said: “Now the challenge will be to show high velocity of sharing of open AI models and datasets (similar to what we're seeing from the labs dominating in open-source AI).”

In October 2025 Laskin said Reflection would release model weights while largely keeping datasets and full training pipelines proprietary. Beam's announcement says the weights, a technical report and the full stack for running and fine-tuning the model are coming, and does not say whether datasets or training pipelines will be released.

Laskin said that Reflection wants builders of open models in the Washington discussion about regulation. He said: “You want multiple voices around the table, both open and closed, and they should be working collaboratively on helping the government regulate AI.”

Laskin says Reflection is working with the US Center for Advancing Innovation and Standards for Super Intelligence and the UK's AI Safety Institute to assess Beam.

Asked whether a model could become too capable to release openly, Antonoglou, a co-founder who co-created AlphaGo, said: “It is possible that you get to a level of capability that you want to just be more careful with how you deploy it.”

Moonshot Releases Kimi K3 Weights as Amodei Rejects Open-Weight Ban

Moonshot AI released the weights for its Kimi K3 model on Monday, a system with 2.8 trillion parameters. Developers can now download, modify and self-host K3, which the report described as the world's

The Implicator

Meta Releases 30B Open-Weight Muse Glimmer and Promises Spark 1.2 Weights

Meta released Muse Glimmer, a 30-billion-parameter open-weight model, Monday. Glimmer distills Muse Spark to run local agents on a single high-end Mac or PC. Meta promised Muse Spark 1.2 weights in co

The Implicator

Germany's Soofi S AI Model Tops All Open-Source Rivals on German Benchmarks

A German research consortium coordinated by the KI Bundesverband released Soofi S, an open-source German-English foundation model, this week, according to its pretraining report. In the team's tests,

The Implicator

Marcus Schuler

San Francisco

Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: editor@implicator.ai