Reflection AI releases Beam, a 501-billion-parameter open-weight reasoning model under Apache 2.0

The news

Reflection AI has released the weights for Beam, the reasoning model it announced earlier in October, under an Apache 2.0 licence. Beam is a sparse mixture-of-experts with roughly 501 billion total parameters and about 23 billion active per token, text-only, with a context window extensible towards a million tokens. The weights are on Hugging Face in BF16, FP8 and NVFP4, alongside a speculative-decoding drafter model, a model card with final benchmark figures, and a safety report; the company also open-sourced its terminal agent harness. Reflection cites day-zero support across the common runtimes, including vLLM, SGLang and Ollama.

Why it matters, and for whom

The headline is the licence. A frontier-scale, US-origin model released under Apache 2.0 (no revenue thresholds, no bespoke terms) is still uncommon, and it is a meaningful addition to the open-weights ecosystem for anyone who needs commercial freedom without reading fine print.

For local use, be clear about the scale. At 501 billion parameters, even a 4-bit quantisation is very roughly 250GB (an approximate figure), so Beam runs on a 512GB unified-memory machine or a multi-GPU server, not a laptop or a single card. Its relevance to local AI is as an option for people with serious hardware, and as a sign of where open, permissively licensed models are going, rather than as a daily local driver. The 23-billion active-parameter count keeps it fast for its size once loaded, but it is the full total, not the active count, that has to fit in memory.

The benchmarks are Reflection’s own

Reflection reports strong figures for Beam, including on coding and agentic tasks, and claims efficiency parity with larger rivals. As with any launch, treat these as the releasing organisation’s own claims until independent evaluations appear; no third-party results were available at the time of writing. Our guidance on reading benchmarks applies here as everywhere.

The history

Beam lands in a steady run of very large open-weight mixture-of-experts models through 2026, following releases such as DeepSeek V4.1 Flash, Tencent’s Hunyuan Hy4, and the still-pending Mistral Large 4. The pattern is consistent: frontier open weights keep arriving, and they increasingly demand datacentre or high-end unified-memory hardware to run, while the models most people can run locally stay in the smaller tiers.

We will add Beam to the model catalogue once we have confirmed the licence file and the exact weight repositories, and can give verified hardware requirements, in line with cataloguing what you can actually run and check.