Reflection AI releases Apache 2.0-licensed weights for Beam, a 501B-parameter model, trained with RL that penalized unnecessary tokens (Matt Marshall/VentureBeat)
Beam activates about 23 billion of its 501 billion parameters for coding, reasoning and agent tasks.
The mixture-of-experts model supports a 1 million-token context window. Its reinforcement learning penalized unnecessary tokens to encourage shorter reasoning. Reflection also released Mirror CLI, the harness used to train it, alongside weights in BF16, FP8 and NVFP4 formats. The company positions Beam as a Western alternative to Chinese open models.
Techmeme's note
The mixture-of-experts model supports a 1 million-token context window. Its reinforcement learning penalized unnecessary tokens to encourage shorter reasoning. Reflection also released Mirror CLI, the harness used to train it, alongside weights in BF16, FP8 and NVFP4 formats. The company positions Beam as a Western alternative to Chinese open models.
Techmeme's note
score 8