Reflection Beam

Reflection Beam is an open-weight large language model from Reflection AI, built for coding, reasoning, and agentic workloads. Its model ID is Beam-1.0-501B-A23B. It is available through a hosted, OpenAI-compatible API and as downloadable weights under the Apache 2.0 license for self-hosted deployment.

Key Capabilities

• Architecture: A sparse mixture-of-experts (MoE) transformer, which activates a subset of its parameters for each token. It has about 501.3 billion total parameters and roughly 23.4 billion active per token. It uses 256 experts across 52 layers (3 dense, 49 MoE), gated grouped-query attention (GQA) combined with sliding-window attention, and a vocabulary of about 152,000 tokens.
• Context length: The model card lists a 1 million-token context.
• Input and language: The model accepts text only. The model card says it was developed predominantly for English.
• Tool use and agentic tasks: The model supports tool integration through the Model Context Protocol (MCP), browsing, and long-horizon multi-step workflows. Reinforcement learning covered coding, agentic, and STEM environments, mostly built through synthetic pipelines.
• Hosted API: The Reflection API exposes chat completions at https://api.reflection.ai/openai/v1, so cURL, any HTTP client, and the official OpenAI SDKs work. Authentication uses an API key from the Reflection platform, read from the REFLECTION_API_KEY environment variable. Responses include a reasoning_content field and token usage that counts reasoning tokens.
• Self-hosting and customization: Weights are published on Hugging Face in Safetensors format in three precisions: BF16, FP8, and NVFP4. The model card states that distillation, fine-tuning, quantization, and merging are not prohibited, and Reflection documents fine-tuning and post-training with open-source training stacks. Reflection states that customers can also deploy in private cloud, on-premises, air-gapped, and edge environments.
• Mirror coding agent: Mirror is a terminal-based coding agent built on Beam, documented as in early beta. It runs interactively, as a one-off prompt with mirror exec, or as an Agent Client Protocol server with mirror acp for editors such as Zed. It provides isolated git worktrees or microVM sandbox workspaces, manual or automatic tool approvals, MCP server connections, extensions and skills, and project instructions read from AGENTS.md. Mirror can also use other models, including OpenRouter, OpenAI, and Anthropic providers and custom endpoints.
• Reported benchmarks: The vendor reports, in BF16, 81.6% on SWE-bench Verified, 97.7% on AIME 2026, 90.6% on GPQA Diamond, 80.1% on Terminal Bench v2.1, and 80.8% on MCP Atlas. For Humanity's Last Exam (HLE), it reports 36.0% without tools and 51.1% with tools. These results were measured at temperature 0.7 and top_p 0.99, and the FP8 and NVFP4 variants score within about two points of BF16 on the same tests.

Audience & Use Cases

• Audience: Developers, researchers, and organizations that need to inspect, customize, self-host, or audit the model they use, including teams with on-premises or air-gapped requirements.
• Use Case: Agentic software engineering, including multi-file coding, debugging, and terminal tasks, through Mirror or through other coding agents configured to use Beam.
• Use Case: Long-context and long-horizon agentic workflows that call tools through MCP.
• Use Case: General assistant-style generation and summarization through the API.
• Use Case: A base model for fine-tuning and post-training on private data.
• Use Case: Code generation from single prompts. The documentation gallery shows examples such as a live subway map built from public data, a fractal explorer, and a browser music sequencer.

Technical Specifications

• License: Apache 2.0, for commercial and research use.
• Formats: BF16, FP8, and NVFP4 weights on Hugging Face. The model card says community quantizations also exist for llama.cpp, Ollama, and LM Studio.
• API compatibility: OpenAI SDK compatible, chat completions request type.