Reflection Releases Beam 501B Open Model for Self-Hosted Coding and Agentic Use
Reflection launched Beam, a 501-billion-parameter mixture-of-experts model with 23B active parameters per token, 1M-token context window, and Apache 2.0 license. Users can…
- Beam ships under Apache 2.0 with downloadable weights, letting teams own, modify, and deploy the full model on compatible GPUs.
- It includes day-0 support in transformers, vLLM, and SGLang plus recipes for supervised fine-tuning and reinforcement learning.
- The sparse MoE design targets long-context reasoning, tool calling, and agentic coding while keeping inference speed closer to a 23B model.
What Beam delivers
On October 16, 2026 Reflection released Beam-1.0-501B-A23B on Hugging Face. The text-only model uses a mixture-of-experts architecture with 501B total parameters but activates only 23B per token. It supports a one-million-token context window and focuses on coding, agentic tasks, and long-context reasoning.
The model includes built-in features for preserving thinking across conversation turns, structured tool calling with a dedicated grammar, and five selectable reasoning effort levels from low to max. Checkpoints are provided in BF16, FP8, and NVFP4 formats. The Apache 2.0 license permits anyone to download, customize, and run the model on their own infrastructure.
Hardware needs and inference support
The BF16 checkpoint is roughly one terabyte. The NVFP4 version reduces this to about 280 GB and runs natively on NVIDIA Blackwell GPUs. Four B200 GPUs or two B300s can hold the weights with room left for long contexts and activations. FP8 checkpoints suit older Hopper and Ampere GPUs.
Day-zero integration is available in Hugging Face transformers, vLLM, and SGLang. A custom Rust-based harness, speculative decoding with a DFlash2 drafter, and expert-parallel loading help maximize performance. Hosted inference options through Hugging Face Inference Providers and Endpoints allow testing without owning hardware.
Post-training recipes and community challenge
The release provides recipes for supervised fine-tuning with TRL using LoRA adapters and for full-parameter or LoRA reinforcement learning with Miles and GRPO. Early tests showed a small LoRA adaptation improved the model's ability to follow user-specified output languages.
Reflection is offering ten thousand dollars in Hugging Face Hub credits for the best community projects built with Beam, including demos on Spaces, quantized versions, synthetic datasets, RL environments, or benchmarks. Vibe evaluations on tool use, codebase retrieval, terminal tasks, math, and structured output are shared to help users explore its strengths.
Why it matters for uncensored AI
Our analysis starts from the source fact that Beam is released under the permissive Apache 2.0 license with fully downloadable weights. This gives developers and organizations the practical choice to own the model, run it on private hardware or cloud instances, and perform supervised fine-tuning or reinforcement learning without depending on any single hosted provider.
For users seeking greater control over agentic tools or private deployments, this expands available frontier-scale options that can be adapted for custom needs. It does not alter the model's built-in behavior or guarantee reduced refusals, as those depend on the base training and any subsequent fine-tuning.
