Reflection AI published open weights for Beam on October 16, 2026 through a Hugging Face Blog post co-authored with the Reflection team. Beam-1.0-501B-A23B is a sparse mixture-of-experts language model with 501 billion total parameters and about 23 billion active per token, released under the Apache 2.0 license for coding and agentic workloads.
The public drop follows Reflection's October 5 preview post, which introduced the architecture, training recipe, and waitlist while promising weights, a technical report, and developer artifacts later in the month. The October 16 materials describe day-0 support in transformers, vLLM, and SGLang, plus BF16, FP8, and NVFP4 checkpoints, a Rust-based harness, tool calling with a dedicated grammar, and reasoning effort settings from low through max.
Architecture notes in the Hugging Face write-up highlight 52 layers, 256 experts with eight routed per token, hybrid sliding-window and global attention, a 1 million token context window, and residual scaling intended to keep activations stable. Reflection's earlier post reported pretraining on 23.8 trillion tokens and a high-compute reinforcement learning campaign on about 10,500 NVIDIA GB300 GPUs that produced more than 100 million rollouts over four weeks.
Benchmark tables published by Reflection and restated on Hugging Face cover agentic coding, reasoning, tool use, and long-context suites. Reflection positions Beam as competitive with larger open models on several coding and agent tasks while emphasizing fewer generated tokens per finished task. VentureBeat independently covered the weight release the same day, quoting Reflection leadership on token efficiency versus Chinese open models and noting Artificial Analysis commentary on early efficiency indicators. Those performance and efficiency figures remain vendor-reported unless separately reproduced.
Deployment guidance stresses that the NVFP4 checkpoint shrinks memory versus BF16 and that practical serving often needs multi-GPU Blackwell or Hopper configurations. Hugging Face lists Inference Providers and Inference Endpoints paths, and Reflection is releasing partner software for self-hosted deployment under Apache 2.0. The Hub model card for reflection/Beam-1.0-501B-A23B is the downloadable artifact accompanying the blog. Users should treat safety claims via Reflection's safety report and third-party red-team mentions as attributed until independent audits are public.
This package treats the open-weight publication as the central event. It is distinct from earlier reporting that Nvidia was in early talks to invest in or acquire Reflection, which concerns ownership rumors rather than a model weight drop.