Overview
Reflection Beam is the first open-weight model from US AI lab Reflection, released on Oct 16, 2026 and positioned for coding and agentic workloads.
It uses a sparse Mixture-of-Experts architecture with 501B total and about 23B active parameters, a 1M-token context window, text-only input and output, and a June 30, 2026 knowledge cutoff.
Beam ships BF16, FP8 and NVFP4 weights under Apache 2.0 for direct download on Hugging Face, alongside a public-beta first-party API and the open-source Mirror CLI terminal coding agent, with day-0 support from more than 40 cloud platforms and inference frameworks.
Founded by former Google DeepMind researchers Misha Laskin and Ioannis Antonoglou and backed by NVIDIA, Reflection frames Beam as the kernel of an open operating system for agents, aimed at teams that want to self-host and customize a frontier model.
Key Features
- 501B MoE architecture: 501B total with ~23B active, 256 experts routing 8 per token for near-23B inference cost
- Apache 2.0 open weights: BF16, FP8 and NVFP4 checkpoints free to download, fine-tune and commercialize with no revenue threshold
- 1M-token context: Built for large codebases, long-horizon tasks and multi-step tool use
- Built-in reasoning: reasoning_effort spans low to max in five steps, thinking before answering
- Open-source Mirror CLI: Official Rust terminal coding agent with isolated workspaces and background sessions, self-hostable
- 40+ day-0 ecosystem: First-day support in Transformers, vLLM, SGLang and Unsloth, plus AWS, GCP and CoreWeave
Use Cases
- Self-hosted coding agents
- Long-horizon agentic tasks and tool use
- Code generation and software engineering
- Private deployment under data compliance
Pros
- Apache 2.0 fully open for commercial use and fine-tuning
- 501B scale with 23B active for efficient inference
- 1M-token long context
- Open-source Mirror CLI and full deployment stack
- 40+ partner day-0 support
Pricing
Weights are free to download under Apache 2.0 (BF16, FP8, NVFP4); the first-party API is a free public beta with no official price yet; OpenRouter lists about $1 input / $3 output per million tokens, with hosts ranging from $0.25/$1 to $1.30/$4.30.
Summary
Reflection Beam is the US open-source camp's answer to Chinese open models like DeepSeek, Kimi and GLM: a 501B MoE fully open under Apache 2.0, paired with the open-source Mirror CLI and 40+ ecosystem partners for self-hostable, customizable coding and agent models. Its SWE-bench Verified 80.9 and AIME 2026 97.8 are vendor-reported, with no independent verification yet.
Version History
- Reflection open-sources Beam (501B MoE) (2026-10-16): Open-sources Beam, its first model, with 501B total, 23B active and a 1M-token context under Apache 2.0 in BF16/FP8/NVFP4; positioned for coding and agents, with vendor-reported SWE-bench Verified 80.9 and AIME 2026 97.8; opens a free public-beta API and the open-source Mirror CLI, with 40+ day-0 partners including AWS, GCP, CoreWeave and OpenRouter; founded by ex-DeepMind researchers Misha Laskin and Ioannis Antonoglou, backed by NVIDIA.