Reflection Beam: 501B Open-Weight MoE Model, Oct 2026

A new Western contender for the open-weight frontier. On October 5, 2026, Reflection introduced Beam, its first open-weight model: a sparse Mixture-of-Experts with 501 billion total parameters, of which 23 billion are active per token, built for coding, reasoning and agentic workloads.

How Beam was trained

• Pretraining on 23.8 trillion curated tokens (web, public sources, licensed datasets), completed in under four weeks on 6,144 NVIDIA GB300 NVL72 GPUs
• A high-compute reinforcement-learning run: 100M+ rollouts on 10.5K GB300 GPUs over four weeks, with rollouts up to 256K tokens and roughly 1.3 billion sandboxes
• Midtraining extends the effective context window to 1M tokens
• A controllable reasoning-effort parameter to trade answer length against accuracy

Benchmarks (vendor-reported)

Reflection lists 80.9 on SWE-Bench Verified, 80.1 on Terminal Bench v2.1, 65.5 on SWE-Bench Pro v1 and 90.5 on GPQA Diamond. The company positions Beam as competitive with GLM 5.2 and approaching Qwen 3.8-Max on coding and agentic tasks, while admitting that Kimi K3 stays ahead on raw capability. Its pitch is efficiency: reasoning scores comparable to GLM-5.2 with 3–4× less inference compute.

Availability

Beam is still in final red-teaming. A preview is open to a select group via a waitlist; Reflection says the weights will ship under Apache 2.0 this month, alongside a technical report, model card and tooling for running and fine-tuning. Beam is text-only.

Rocks take

For anyone running local AI next to Home Assistant, a 501B MoE is datacenter-class, not Raspberry-class—but an Apache 2.0 licence with 23B active parameters makes it a serious candidate for self-hosted coding agents on multi-GPU rigs. Until the weights and the technical report are public, treat every benchmark above as a company claim.

FAQ

Can I download Beam today? Not yet—only waitlist access; weights are promised later in October 2026.

Is it multimodal? No, Beam is text-only.