Reflection AI unveils Beam, a 501B open-weight model

Reflection AI unveiled Beam, a 501B open-weight MoE with 23B active parameters, claiming lower-compute coding performance, with Apache 2.0 weights due this month.

Reflection AI introduced Beam, its first open-weight model: a text-only sparse mixture-of-experts with 501 billion total parameters and 23 billion active, aimed at coding, reasoning, and agentic work, with Latent Space also citing scientific work. It was pretrained on 23.8 trillion tokens—Hacker News says web and licensed data, while Latent Space says training was from scratch in the US and included OCR over hundreds of millions of PDFs—and TechCrunch reports a 1 million-token context window. Reinforcement learning used about 10,500 NVIDIA GB300 GPUs and more than 100 million rollouts over roughly four weeks; Help Net Security says scores were still rising when the run stopped, and The Decoder says an early version is available to select users. Performance claims are disputed in detail: the company says Beam matches GLM 5.2 on advanced reasoning at roughly three to four times less inference compute and approaches Qwen 3.8-Max (also styled Qwen3.8-Max) on coding, with reported scores of 80.9 on SWE-bench Verified and 80.1 on Terminal-Bench v2.1, but Help Net Security says the compute figure is an estimate from active parameters and generated tokens rather than measured serving cost, notes Terminal-Bench trails GLM 5.2 (81.0), Kimi K3 (88.3), and DeepSeek V4.1 Flash (90.6), and reports DeepSWE v1.1 at 44.4; TechCrunch says the claims are not independently verified, and Latent Space observers place Beam near GLM-5.2 and behind DeepSeek V4 and GLM 5.3. Apache 2.0 weights, a technical report, and a model card are planned later in October 2026 after final red-teaming, including distribution through hyperscalers, neoclouds, and open-source libraries. Funding accounts disagree: TechCrunch says former Google DeepMind researchers raised about $4.7 billion at a $25 billion pre-money valuation and signed more than $7 billion in compute deals with SpaceX and Nebius, while The Decoder says the ex-DeepMind founders raised $2 billion at an $8 billion valuation.

• Beam is Reflection AI’s first open-weight, text-only sparse mixture-of-experts model: 501 billion total parameters and 23 billion active.
• It was pretrained on 23.8 trillion tokens; TechCrunch reports a 1 million-token context window.
• Reinforcement learning used about 10,500 NVIDIA GB300 GPUs and more than 100 million rollouts over roughly four weeks.
• Reported scores include 80.9 on SWE-bench Verified, 80.1 on Terminal-Bench v2.1, and 44.4 on DeepSWE v1.1.
• On Terminal-Bench v2.1, Beam trails GLM 5.2 (81.0), Kimi K3 (88.3), and DeepSeek V4.1 Flash (90.6).
• Reflection claims GLM 5.2-level reasoning at about 3–4× less inference compute; Help Net Security calls that an estimate, not measured serving cost, and TechCrunch says the claims are unverified.
• Apache 2.0 weights are planned later in October 2026 after final red-teaming; The Decoder says an early version is available to select users.
• Funding figures conflict: TechCrunch reports about $4.7 billion raised at a $25 billion pre-money valuation, while The Decoder reports $2 billion at an $8 billion valuation.