Skip to content
NewsModel release

Reflection AI previews Beam, a 501-billion-parameter model aimed at China's open-weight leaders

· by Pondero Newsdesk

The short version

Reflection AI unveiled Beam on October 5 with benchmark claims against Z.ai's GLM-5.2, but the Apache 2.0 weights will not post until later this month and only a waitlist is open today.

Reflection AI previews Beam, a 501-billion-parameter model aimed at China's open-weight leaders

Reflection AI, a Brooklyn startup valued at $25 billion in its most recent funding round, unveiled a 501-billion-parameter model on October 5 that it says matches a leading Chinese model's reasoning scores while using three to four times less inference compute. Nobody outside Reflection can actually run it yet. The company opened a waitlist for early access the same day, and the Apache 2.0 weights will not post until later this month, so anyone sizing Beam up against GLM-5.2 or Inkling today is working from Reflection's own numbers alone, with no independent rerun to check them against.

What Reflection showed

Beam is a sparse mixture-of-experts model with 501 billion total parameters and 23 billion active per token, pretrained on 23.8 trillion tokens with an effective context window of 1 million tokens, per Reflection's launch post. It is text-only, built for coding, reasoning, and long-running agent work rather than images or audio. Reflection says Beam scores comparable to Z.ai's GLM-5.2, a roughly 744-billion-parameter model with 40 billion active parameters, on advanced reasoning benchmarks while using three to four times less inference compute, and that it beats the leading Western open models it tested against, according to TechCrunch's report on the launch. Those performance claims have not been independently verified. The same benchmark tables show Beam ahead of Inkling, the open model Thinking Machines Lab released in July, on four coding tests both companies report, though Inkling is multimodal and Beam is text-only.

On Reflection's own benchmark tables, Beam scored 80.9 on SWE-Bench Verified, 90.5 on GPQA Diamond, and 97.8 on AIME 2026, per the launch post. Those numbers come from a training run Reflection says used 10,500 Nvidia GB300 GPUs over four weeks to generate more than 100 million reinforcement-learning rollouts, with a separate pretraining run on 6,144 GB300 GPUs finishing in under four weeks. The company credits that reinforcement-learning investment, not just model size, for Beam's efficiency claim, saying performance kept improving as it added RL compute with no sign the gains had topped out.

The blog post frames this as a preview, not a ship date. Reflection wrote that Beam "is undergoing final red-teaming and evaluations," pointed developers to a sign-up page for early access, and said it would release the weights, a technical report, a model card, and developer tooling "later this month" under an Apache 2.0 license, with distribution partners lined up across hyperscalers and neoclouds.

Enterprises can join the waitlist now; self-hosting waits on the weights

Reflection is pitching Beam at enterprises, governments, and what the company calls sovereign AI buyers who want to train the model on their own proprietary data rather than route requests through a closed API. Nvidia CEO Jensen Huang has pushed that "AI factory" framing for years, and Nvidia is one of Reflection's backers, per TechCrunch. The company has already started testing the idea with South Korea's Shinsegae Group on a sovereign AI factory partnership, the same report says. That pitch depends on a buyer being able to download and run the model inside its own infrastructure, which is exactly what is not possible yet. Every benchmark number Reflection has published so far was produced on Reflection's own hosted checkpoint, not on a build a customer can inspect or reproduce. A team evaluating Beam against GLM-5.2, Inkling, or another open model for an actual deployment has reason to treat this week's numbers as a claim to revisit once the weights, the technical report, and any third-party reruns are public, rather than as a result to plan around today.

Context and reactions

Reflection was founded in 2024 by two former Google DeepMind researchers and has raised roughly $4.7 billion from backers including Nvidia, Sequoia Capital, and Lightspeed Venture Partners, per TechCrunch's reporting, which cites PitchBook data. Its most recent funding round valued the company at $25 billion pre-money, the same report says. This summer, Reflection signed compute deals worth more than $7 billion combined with SpaceX and Nebius to lock in access to Nvidia's GB300 chips through 2029. That level of compute spending points to how Reflection sees the field: it names Anthropic and OpenAI as closed-model rivals, DeepSeek, Qwen, and Z.ai as the Chinese open-weight labs setting the pace it wants to match, and Mistral, Meta, and Cohere among the Western players it says it is also racing. TechCrunch reported that Reflection's launch confirmed weekend reporting from Axios that a release was close, and that Reflection did not respond to TechCrunch's request for comment before publication.

What to watch next

Two things will determine whether Beam's debut holds up. Independent evaluators such as Artificial Analysis or LMSYS have not yet run their own numbers on Beam, so the GLM-5.2 comparison rests entirely on Reflection's own report until that happens. And the "later this month" timeline for the actual weights, technical report, and model card is itself unconfirmed; Reflection has not set a specific date. Whether Reflection ships on that schedule, and whether its benchmark claims survive outside testing, are the facts that turn this preview into a release.

Sources