Reflection AI has released Beam, a 501 billion parameter language model, and put the weights out in the open.
Reflection AI published Beam in a blog post on October 5. The company is releasing it as open-weight, meaning the trained parameters are downloadable rather than restricted to a paid API like the flagship models from OpenAI or Anthropic. That makes Beam one of the largest openly available models released so far. The post quickly surfaced on Hacker News, picking up 93 points and 24 comments.
At 501 billion parameters, Beam sits in the same size class as DeepSeek's V3 and well past Meta's 405 billion parameter Llama 3.1, another sign that the open-weight frontier keeps scaling up rather than down. Size alone settles nothing, though - without independent benchmarks, a parameter count is just a number, not a verdict on how Beam actually performs.
For context, 24 comments on a 93 point Hacker News post is modest chatter for a release this size - the kind of muted reaction that suggests the jury is still out.