Reflection AI releases 501-billion-parameter Beam weights under Apache 2.0

Summary

• Reflection AI published open weights for Beam, its 501-billion-parameter agentic coding model, on Hugging Face.
• The Apache 2.0 license permits commercial use and modification, with day-one support across major inference frameworks.
• Enterprises can now run a frontier-scale agent model on private, sovereign or air-gapped infrastructure without licensing fees.

The latest

Open weights for Beam, a 501-billion-parameter sparse mixture-of-experts model with 23 billion active parameters, went live on Hugging Face on 16 October 2026 under an Apache 2.0 license. Reflection AI said the model is built for coding, reasoning and long-horizon agentic workloads, and released it in BF16, FP8 and NVFP4 precisions alongside a draft model for speculative decoding.

Details

• The architecture: Reflection AI described Beam as a sparse mixture-of-experts system with 501 billion total parameters and 23 billion active at inference. The company shipped optimized serving configurations and fused kernels developed with partners to cut latency and raise throughput, and said lower precisions reduce memory requirements.
• The license: Apache 2.0 allows commercial deployment, modification and redistribution without licensing fees. Reflection AI framed the release around a stated goal of making intelligence open and accessible to all, and a position that the model is the kernel while everything above the infrastructure layer stays open.
• Training scale: The company reported roughly 1.3 billion isolated sandboxes used across reinforcement learning, evaluation and experiments, with a peak fleet of 883,006 vCPUs spanning 23 clusters on Google Cloud and Amazon Web Services. It said it achieved near-zero cluster downtime. Training data was not disclosed.
• Earlier figures: According to Unite.ai, citing Reflection materials, Beam was first introduced on 5 October 2026 with early access through a waitlist, carries a context window of up to 1 million tokens, and was pre-trained on 23.8 trillion tokens. A reinforcement learning run produced more than 100 million rollouts on 10,500 NVIDIA GB300 processors over four weeks.
• A second account: SiliconANGLE reported the model was trained over two months on several thousand high-performance GPUs and approaches the performance of models up to 4 times larger on some benchmarks.
• The agent harness: Reflection AI open-sourced Mirror CLI, a terminal harness trained jointly with the model, pairing a Python interface with a Rust agent core. Beam was also tested with Pi from Earendil, OpenCode from Anomaly and Hermes from Nous Research, and the company said more harness support is coming.
• Customization: Fine-tuning and reinforcement learning recipes were published with NVIDIA, covering NeMo RL, AutoModel and Megatron, alongside RadixArk, Hugging Face TRL, Anyscale SkyRL and Axolotl. Self-distillation recipes, generating traces then training LoRA adapters on verified ones, were promised within days.
• Where it runs: Day-one inference and training support covers Transformers, vLLM, SGLang, TokenSpeed, NVIDIA Dynamo, Unsloth and Axolotl. Cloud and hosted availability spans more than 20 providers including AWS, Microsoft Azure, Google Cloud, Databricks, CoreWeave, Cloudflare, Oracle and Together AI, with on-premise and air-gapped deployment through Dell and HPE.
• Pricing: The Reflection platform at platform.reflection.ai opened as a free public beta. No pricing beyond the free period was announced, and no minimum hardware requirements were published, though Reflection said its production board runs Beam on SGLang over NVIDIA Blackwell.
• Self-reported demos: Reflection said Beam built a mobile snake game in 2:19.7 against 4:56.4 for a comparison model, and that a voxel pagoda demo needed a single line of repair. The company noted some demonstrations ran through stealth previews.

Between the lines

The release pairs an unrestricted license with on-premise and air-gapped distribution through Dell and HPE, which puts a frontier-scale agent model inside environments that cannot send data to a hosted API. The jointly trained Mirror CLI and the published fine-tuning recipes suggest Reflection is competing on the surrounding toolchain rather than on model access alone, consistent with its stated position that the model is the kernel.

What's next

Watch for the self-distillation recipes Reflection said would arrive within days, the final benchmark numbers in the model card, the separate Beam safety technical report, and whether pricing follows the end of the free public beta.

Source: Reflection AI, Unite.ai, SiliconANGLE