Reflection AI · Language model

Beam 1.0

Reflection AI's first model: a 501-billion-parameter mixture-of-experts with 23 billion active and a 1M-token context window. Weights went up on Hugging Face under Apache 2.0 on October 16.

MakerReflection AI
ReleasedOctober 16, 2026
API price (per 1M tokens)Not published
Open weightsYes

Best for

  • Teams that want a US-made model under Apache 2.0
  • Fine-tuning and research (post-training recipes are published)
  • Self-hosting on data-centre hardware

Strengths

  • Apache 2.0 license
  • Second among US open-weight models on the Artificial Analysis index, behind Muse Spark Open
  • Token-efficient: about 26K output tokens per Artificial Analysis index task

Watch out for

  • Scores 30 on the Artificial Analysis index, well behind GLM-5.3 and DeepSeek V4.1 Flash
  • No public API yet, so you have to host it yourself
  • Reflection's own coding scores haven't been reproduced independently

Ranking history

EditionPower RankingsCodingBest ValueOpen-Weight WatchImage GenerationVideo Generation
Week 45, 2026———#6 (79)——
Week 44, 2026———#5 (80)——