At a Glance
| Vendor | Reflection AI |
| Family | Beam |
| Launched | Reflection announced Beam on October 5, 2026, with weights reported on Hugging Face on October 16 as reflection/Beam-1.0-501B-A23B. It is a US-built open-weight model aimed at coding, reasoning, and agent work. It is text-only. |
| Context window | 1M tokens on Reflection's API, with 128K maximum output. The model has 501 billion total parameters and 23 billion active per token, across 52 layers. |
| Pricing | As of October 27, OpenRouter listed Nebius at $0.25 / $1.00 per million input and output tokens, and Baseten, Together, and DigitalOcean at $0.40 / $1.20. Reflection's own API is in a free-to-start public beta. |
| Access channels | Hugging Face weights, Reflection's API, OpenRouter, and hosts including Nebius, Baseten, Together, and DigitalOcean. |
Notable Benchmarks
Reflection reports 80.9 on SWE-bench Verified, 80.1 on Terminal-Bench 2.1, 97.8 on AIME 2026, and 90.5 on GPQA Diamond. It says Beam matches GLM-5.2 on advanced reasoning using 3 to 4 times less inference compute, and it concedes that Kimi K3 is ahead on raw capability. The compute figure is an estimate that leaves out serving overhead, and TechCrunch noted the scores had not been independently verified.
Strengths
An Apache 2.0 license, unusually open at this size, plus efficient inference and very low hosted prices. Reflection says it does not prohibit distillation or training on outputs.
Limitations
Server hardware is still needed. The vLLM recipe calls for about 353 GB of GPU memory at NVFP4 (4 B200 or B300, or 8 H100), and 1,204 GB at BF16. Kernels are CUDA-only for now. There is no native image input. The Financial Times reported on October 10 that Nvidia was in talks to expand its investment or acquire the company, and neither has confirmed a deal.
Brand-Visibility Implications
Low hosted prices make Beam a plausible default inside third-party agents and research tools, where brands will not see which model is behind an answer. Open weights also mean behaviour can vary by host and any fine-tuning. Treat it as a monitoring item for developer-facing categories, not a new consumer surface. See Muse Spark Open hardware requirements and how much RAM each open-weight model needs.
How Presenc AI Tracks This Model
Presenc AI monitors brand visibility on Reflection AI's Beam as part of continuous multi-platform AI visibility tracking. We sample Reflection Beam across representative prompt sets daily, compare against competitor performance on the same prompts, and flag material mention-rate changes so brand teams can respond quickly when AI representation shifts.