Reflection Beam
Reflection's first open model: a 501B text mixture-of-experts under Apache 2.0, built to be efficient at inference.
In short. Beam is the first open model from the US lab Reflection AI. It only handles text, it is strong at coding and reasoning for its size, and its Apache 2.0 license lets you use and host it freely. It is large, so you need a multi-GPU server.
Summary
Reflection AI announced Beam on 5 October 2026 as its first open-weight model and published the weights on Hugging Face under the Apache 2.0 license on 16 October 2026.
Beam is a text mixture-of-experts model with 501 billion parameters in total and 23.4 billion active per token. It has 52 layers (3 dense layers, then 49 mixture-of-experts layers), 256 experts with 8 routed per token and no shared expert. Attention is gated grouped-query attention, and 43 of the 52 layers use a sliding window of 512 tokens, so only 9 layers keep a cache that grows with the whole context. That is how a 1-million-token window stays affordable. The knowledge cutoff is 30 June 2026. Reflection says pretraining used 23.8 trillion tokens.
Three checkpoints are published: BF16 (about 1 TB), FP8 (about 0.5 TB) and NVFP4 (about 0.3 TB), plus a speculative-decoding drafter. Hugging Face suggests four B200 GPUs for the NVFP4 version.
Reflection's own framing is modest: it says frontier open models such as Kimi K3 remain ahead on raw capability, and that Beam's advantage is lower inference cost, with scores comparable to GLM-5.2 on reasoning benchmarks. These are vendor claims, and benchmark numbers changed between the launch blog and the final model card, so check the card.
Good for
- Companies that want a US-built open model with Apache 2.0 terms and no revenue conditions
- Coding and reasoning agents where inference cost matters
- Distillation, fine-tuning and quantization work; the model card explicitly allows them
- Long-context text workloads up to 1 million tokens
Watch out for
- It handles text only, with no image input
- Even the NVFP4 build is about 294 GB, so a multi-GPU server is required
- Reflection's efficiency and benchmark comparisons are self-reported
- Some larger open models, such as DeepSeek V4.1 Flash, score higher on the coding benchmarks in Reflection's own tables
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Estimates for the Beam 1.0 NVFP4 variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.
| Variants | Parameters | BF16 | FP8 / INT8 | 4-bit (Q4) | Fits on |
|---|---|---|---|---|---|
| Beam 1.0 BF16 | 501B 23.4B active | 1.15 TB | 576 GB | 346 GB | Big-memory desktop |
| Beam 1.0 FP8 FP8 | 501B 23.4B active | 1.15 TB | 576 GB | 346 GB | Big-memory desktop |
| Beam 1.0 NVFP4 NVFP4 | 501B 23.4B active | 1.15 TB | 576 GB | 323 GB | Big-memory desktop |
Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.
The license in plain words
One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.
See also
- Best open-source LLMs in 2026
- Ripple-65B Preview: a dense 65B agent model under Apache 2.0
- Giant open models now ship in 4-bit: what it means for memory planning
Frequently asked questions
- What is Reflection Beam?
- Beam is the first open model from the US lab Reflection AI. It only handles text, it is strong at coding and reasoning for its size, and its Apache 2.0 license lets you use and host it freely. It is large, so you need a multi-GPU server.
- Can I use Reflection Beam commercially?
- Yes. Reflection Beam is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is Reflection Beam open source?
- Reflection Beam is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
- Can I run Reflection Beam on my own computer?
- Only on a machine with a lot of memory. The smallest variant, Beam 1.0 NVFP4, needs about 323 GB at 4-bit precision. Desktops with 128 to 512 GB of unified memory can hold it, at modest speed.
- When was Reflection Beam released?
- Reflection Beam was released by Reflection AI on October 16, 2026.
- How long a context window does Reflection Beam have?
- Reflection Beam supports a context window of up to 1M tokens.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- Introducing Beam: Reflection's 501B open-weight model (Reflection, 5 Oct 2026) reflection.ai
- Beam is now available (Reflection, 16 Oct 2026) reflection.ai
- Beam-1.0-501B-A23B model card (Hugging Face) huggingface.co
- Welcome Beam: a 501B open model you can own, customize, and scale (Hugging Face blog) huggingface.co
Release date and facts last verified November 8, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| MiniMax M3.1 Flash MiniMax | 428B 23B active | 1M | MiniMax Community (Conditional) | 277 GB Big-memory desktop | |
| DeepSeek V4.1 Flash DeepSeek | 552B 16B active | 1M | MIT (Permissive) | 561 GB GPU server | |
| Inkling Thinking Machines Lab | 276B – 975B 41B active | 1M | Apache 2.0 (Permissive) | 188 GB Big-memory desktop | |
| Ripple-65B Preview Ricursive Intelligence | 64.8B | 1M | Apache 2.0 (Permissive) | 45 GB Workstation |