Reflection has released Beam, an open LLM that shifts attention from raw parameter count to the cost of reasoning. The spec sheet lists 501 billion total parameters and 23 billion active parameters, a 1 million token context window, and three inference numerics: BF16, FP8, and NVFP4. The model is instruction-tuned and developed primarily for English.

The most relevant point is not overall size, but the engineering thesis Reflection is pushing: Beam is designed to use about one third of the reasoning tokens of leading open models at comparable quality. In a context where operating cost grows with generated tokens, that choice translates into a direct TCO benefit for teams running on-premise workloads. Fewer reasoning tokens mean less pressure on latency and compute resources, especially for long coding, debugging, or tool-use tasks.

But there is a catch that the label '501B MoE' should not hide. The 23 billion active parameters reduce per-token compute, because only part of the experts is involved at each step; however, the full set of experts must stay available in memory. With 256 experts and 52 layers — 3 dense and 49 MoE — the VRAM pressure is not that of a 23 billion parameter model. It is that of a 501 billion parameter model, mitigated by quantization. FP8 and NVFP4 exist precisely for this: reducing weight footprint and increasing throughput. In return, they shift attention to compression quality, which becomes as important a deployment variable as the architecture itself.

The published benchmarks show a strong profile on reasoning and coding tasks. Beam in BF16 reaches 97.7% on AIME 2026, 90.6% on GPQA Diamond, and 81.6% on SWE-bench Verified. Differences among BF16, FP8, and NVFP4 remain contained on many tests, a useful signal for teams wanting to exploit lower-precision formats without losing too much. Tool calling and browsing results are more uneven: MCP Atlas exceeds 80%, while AutomationBench stops at 39.3% in BF16. That spread describes a model tuned for coding, scientific reasoning, and tool orchestration, not a uniform generalist.

For those watching on-premise deployment, Beam sharpens a distinction that is often blurred: inference compute cost drops, but the capital threshold to bring the model in-house remains high. The choice between BF16, FP8, and NVFP4 is not purely technical, because it determines which accelerators can serve the model and what TCO is realistic. The numbers show that compression holds on several benchmarks, but they do not say how it behaves on heterogeneous hardware or when context really approaches one million tokens. That is exactly why Beam is an interesting test: it promises to reason with fewer tokens, but the memory bill is not discounted.