Reflection: Beam 1.0 501B A23B

Reflection's Beam 1.0 501B A23B demonstrates strong all-around capabilities, leveraging a sparse mixture-of-experts architecture paired with a massive 1-million-token context window. Across comprehensive benchmarking, the model demonstrates competitive response times and offers competitive pricing. It also exhibits exceptional reliability, maintaining a 98% success rate with minimal technical failures. Performance across categories reveals distinct strengths, particularly in structured execution. The model excels in instruction following, scoring 82.0% accuracy to place in the 84th percentile, proving highly adept at adhering to multi-step constraints and strict formats. It also shows notable strength in coding, securing 94.0% accuracy (80th percentile) across various programming languages. While raw scores in mathematics (92.0%) and complex reasoning (92.0%) remain high, they land closer to the industry median relative to peer models. A notable weakness is its latency during extended reasoning tasks, where response times slowed considerably into the 18th percentile. Overall, Beam 1.0 delivers dependable execution, robust coding competency, and exceptional prompt adherence, making it a reliable and cost-effective engine for complex, long-horizon workflows.