Evidence at a glance
The mechanism in one line
Compress the visual or contextual input before the main reasoning path.
Route or verify the expensive step instead of repeating the full path.
Translate the mechanism into a bounded deployment or evaluation check.
After Open Weights, Ownership Gets Concrete
On October 16, 2026, Reflection released Beam on Hugging Face as the new lab’s first open-weight model. It targets coding, reasoning, and agentic tasks under the Apache 2.0 license, and teams can run and customize it on their own hardware, infrastructure, or cloud. The release guide also provides inference, supervised fine-tuning, and reinforcement-learning recipes, alongside deployment paths through Transformers, vLLM, and SGLang.
The release matters not simply because the model has 501 billion parameters, but because “owning a model” breaks down into distinct promises: receiving weights, being able to deploy them, and being able to continue training do not mean owning the training data or the full training process. Beam gives users practical paths for the first parts of that equation, while leaving hardware and serving costs to them. For a technical leader, the question is not whether the model is open enough, but whether the team can turn that openness into a sustainable operating and customization capability.
23B Active Parameters Do Not Mean a 23B-Scale Deployment
Beam is a sparse mixture-of-experts model: it has 501 billion parameters in total, but activates about 23 billion for each token. It has 52 layers, with dense layers at the start and 49 MoE layers after them; each MoE layer has 256 experts, and each token is routed to eight. This reduces the amount of computation used for an individual token, but the other experts’ weights do not disappear.
Deployment requires keeping two ledgers separate. The compute ledger asks how many experts are used per token; the storage ledger asks how many weights must be available at runtime. The guide puts the BF16 checkpoint at about 1 TB and the NVFP4 version at about 280 GB. NVFP4 is not a lightweight option for every system: the guide positions it for NVIDIA Blackwell, while suggesting FP8 for earlier GPU generations. “23B active” helps explain sparse computation, but cannot by itself estimate memory, hardware requirements, or total cost.
A Million-Token Context Is Both an Architectural Choice and a Cache Bill
Beam combines local and global attention in the same network. Forty-three layers use a sliding window over the previous 512 tokens, while nine handle the full context; the release guide says this arrangement reduces the growth in KV-cache overhead as sequences get longer. The model supports up to a one-million-token context, but that limit describes what it can accept, not a cost-free way to process long inputs.
The guide’s estimates show the difference: for a single BF16 sequence at one million tokens, the KV cache is about 36.08 GiB; if all 52 layers used global attention, the estimate would be about 208 GiB. Beam is therefore not simply stretching the window; it trades off long-range information and resource use through different layer roles. In practice, whether long context is worthwhile depends on whether a request needs that much history and whether the serving system can absorb the cache. The maximum context length alone does not answer that question.
Customization Paths Are Not a Zero-Barrier Capability
Beam’s release offers more than weights. The guide lists BF16, FP8, and NVFP4 checkpoints, examples for running the model with Transformers, vLLM, and SGLang, and recipes for SFT, LoRA, and RL. Reflection also says full-parameter RL is supported with Miles. These resources give teams a path from inference toward fine-tuning or reinforcement learning without requiring them to build every workflow from scratch.
But recipes and code entry points answer where to start; they do not solve compute, engineering, or data constraints. Reflection says its pretraining used 23.8 trillion tokens and that its reinforcement-learning run used 10,500 NVIDIA GB300 GPUs for four weeks, generating more than 100 million rollouts. Those figures are company-reported, and they underline that reproducing the training effort is entirely different from downloading the weights. For most teams, a more practical customization path may be to test inference and smaller-scale post-training first, then decide whether heavier training investment is justified.
Benchmarks and Openness Both Leave Questions to Check
The published scores should not be compressed into a single unconditional verdict. Reflection’s introduction lists 80.9 on SWE-bench Verified, while the Hugging Face guide’s precision table reports 81.6% for BF16, 80.2% for FP8, and 80.7% for NVFP4; the available material does not explain the difference. In Reflection’s own Terminal Bench v2.1 table, Beam scores 80.1, below the listed results for GLM 5.3, Kimi K3, Qwen 3.8 Max, and DeepSeek V4.1 Flash. One score cannot stand in for its coding and agentic performance as a whole.
The scope of openness has a similar boundary. The Apache 2.0 license specifies the terms for using the model weights, but Reflection says its pretraining data includes proprietary licensed data, and the public pages do not provide the complete dataset. The release guide also does not establish that a full safety evaluation has been made public. Beam is best understood as an open-weight model that can be deployed and further adapted, not as a system with fully transparent data, training, and safety evidence. Before adoption, teams should make task-specific testing, target-hardware costs, and required data-governance evidence conditions of deployment.