Open-weight models keep climbing, and the questions get harder

Reflection's Beam is a sparse Mixture-of-Experts model with 501B total parameters and 23B active, pitched at coding, reasoning, and agentic work. Commenters welcomed another open-weight release but immediately asked where the data comes from. One joked about inheriting a data center full of GPUs and free electricity and still not knowing where to get the data. Another did a double-take at a demo caption about recreating a viral X puzzle. Related threads: a Dust paper on pretraining transformers without backpropagation, where a 243M model beat one 120x smaller, and an older Ask HN where people compare Qwen 3.6 27B to Claude Haiku 4.5 for daily coding.

The pattern: the open ecosystem is no longer just trying to catch up on benchmarks. It is attacking the cost structure. Dust is less efficient than backprop but easier to parallelize, and commenters immediately floated hybrids that fine-tune a backpropped checkpoint. Local-model people still complain about tokens per second against the cloud, but the gap is described in tradeoffs now, not impossibility.

The counterpoint is credibility. Demo captions, training data provenance, and 'it's about as good as Haiku' claims are all getting scrutinized. Openness is table stakes. Trust in how the model was built is the new fight.