Reflection AI’s Beam 1.0 is out, with AC2 inference and post-training support
Applied Compute says 40 steps of AC2 post-training raised Beam’s score on a long-horizon coding task from 0.07 to 0.215.
TLDR
Applied Compute says Reflection AI’s Beam 1.0 was released on October 16 under the Apache 2.0 license. It describes Beam as a 501-billion-parameter mixture-of-experts model with 23 billion active parameters and says AC2 offers day-zero inference and post-training support. Applied Compute reports that a brief reinforcement-learning post-training run in AC2 tripled Beam’s score on a challenging long-horizon agentic coding task in 40 steps, from 0.07 to 0.215.
Combined views
—
1 Source, first seen ago
