Reflection releases Beam, its first open-weight model, under the Apache 2.0 license

Reflection released Beam, its first open-weight model, on Oct 16. The Hugging Face card lists 501.3B total parameters, 23.4B active, a 1M token context window, text in and text out, a June 30, 2026 knowledge cutoff and the Apache 2.0 license. Reflection also released FP8 and NVFP4 weights, a DFlash2 drafter and Mirror CLI, the open source coding agent tool it trained Beam on. Downloads are about 1,003 GB at BF16, 503 GB at FP8 and 294 GB at NVFP4. Arena ranks it #12 among open models in Agent Arena and #22 among open models in Code Arena WebDev. Reflection says Beam uses "3x fewer reasoning tokens"; that is Reflection's own claim.
How you’d use it: Download Beam from Hugging Face; the NVFP4 weights are about 294 GB.