Reflection AI Launches Beam: 501B-Parameter Open Model Takes on DeepSeek and Qwen

What It Is First

Beam is the first model Reflection AI has ever released. The startup was founded in 2024 by former Google DeepMind researchers Misha Laskin and Ioannis Antonoglou, backed by Nvidia, Sequoia, and Lightspeed, with a reported valuation around $25 billion in its last round.

The model itself is a text-only MoE (mixture of experts): 501B total parameters, 23B active per token, a 1M-token context window. Per the company’s technical note, Beam pre-trained on 23.8 trillion tokens, then ran RL on 10,500 Nvidia GB300 GPUs for four weeks, producing over 100 million rollouts.

The targets are named explicitly: DeepSeek and Moonshot’s Kimi (Kimi K3), with Beam claimed close to Zhipu’s GLM-5.2 on coding and agentic tasks, approaching Qwen3.8-Max. The audience: enterprises and governments wanting locally deployed, customizable open-weight models.

What the Benchmarks Say

Reflection published two numbers itself. On Terminal-Bench 2.1, Beam scored 80.1 versus GLM-5.2’s 81.0; on SWE-bench Pro v1, Beam hit 65.5 versus GLM-5.2’s 62.1. The company claims GLM-5.2-level reasoning at one-quarter to one-third the inference compute.

Three caveats. First, these are company-published scores, not yet independently verified, as TechCrunch noted. Second, the comparison is against GLM-5.2, not the newer GLM-5.3 — in Reflection’s own tables, GLM-5.3 and Kimi K3 rank ahead of Beam on most tests. Third, “cheaper compute” refers to theoretical FLOPs, excluding real-deployment costs like prefill and long-context attention overhead; true costs await real testing.

For reference: GLM-5.2 has ~744B total parameters with 40B active per token. Beam catches it with fewer active parameters by piling on RL — 10,500 GB300s for four weeks, a training scale genuinely rare in the open camp. The money: reported compute commitments exceed $7 billion, including a billion-dollar-scale Nebius partnership and ~$150M/month renting SpaceX’s Colossus 2 cluster.

When Can You Use It

Beam isn’t downloadable yet. The company says weights will drop later in October under Apache 2.0, alongside a technical report and model card. You can join the waitlist now for an early build; final red-teaming and evals are underway.

Apache 2.0 means commercial use, modification, and redistribution are all allowed — good news.

Why It Matters

In open weights, DeepSeek and Qwen have led all year: Vercel AI Gateway token data shows open models hitting a 62% share record one Saturday. The West has lacked a credible answer. Beam is the first Western player to challenge Chinese open models head-on, with training scale and capital backing to match.

CEO Laskin’s line: closed models are like renting; an “ownership market” is forming. The business model hides in that sentence: open weights are the wedge — the real money is building “AI factories” for enterprises and nations, customized on customer data, deployed locally, billed for compute and services.

The suspense resolves at the end of October: once weights drop, can independent evals reproduce the company’s numbers? Then we’ll see whether Beam truly matches up — or it’s benchmarketing.