Hacker News Reader

So marginally better (and occasionally worse) performance for an order of magnitude larger training costs…?

14b model performs comparably with 32b size. the improvement is huge

are we only comparing them in terms of text completion accuracy? does it also improve perf on benchmarks?