Reflection AI releases open-weight Beam coding model with 501B total / 23B active parameters under Apache 2.0

On October 16, AI startup Reflection AI released BF16, FP8, and NVFP4 weights of Beam, its first open-weight model, on Hugging Face under Apache 2.0. Beam is a sparse mixture-of-experts model with 501B total parameters and ~23B active per token, text generation only, with a ~1M token context window and a knowledge cutoff of June 30, 2026.

Reflection was founded by former Google DeepMind researchers Misha Laskin and Ioannis Antonoglou and is backed by Nvidia. The company open-sourced Mirror CLI, the Rust-based command-line agent framework used to train and test Beam, and announced integrations with more than 40 cloud and developer tooling partners. On OpenRouter, hosted Beam 1.0 lists at $1 input / $3 output per 1M tokens, with cache reads at $0.10.

Google Cloud announced Day 0 support for Beam on the same day: up to 5x throughput and 8x latency improvements on A4X Max (NVIDIA GB300 NVL72) instances via llm-d (vLLM) and NVIDIA Dynamo (SGLang), and up to 2x throughput on G4 VMs (NVIDIA RTX PRO 6000).

Reflection's own evaluation reports BF16 weights at 80.1% on Terminal-Bench 2.1 and 81.6% on SWE-bench Verified. These are vendor-reported figures and have not been independently reproduced.