I'm BULLISH on @_inception_ai. We ran Mercury 3 through our open-ended coding benchmarks. They're the only lab making serious progress at scaling diffusion architectures for difficult text and coding work. Their latest release, Mercury 3, is actually competitive and not far from the Pareto front, despite a paltry amount of funding compared to other neolabs - the only competitive dLLM in general text-based reasoning today. Benchmark Results: - The hype is real -- it's a competitive model at its price point. We ranked it slightly lower than its press release, roughly equivalent to GPT-5.6 Luna and Muse Spark 1.3 average across coding tasks due to some errors we discuss below. - It uses more reasoning tokens on hard problems than I expected from a diffusion architecture, so its speed is not a huge differentiator over GPT-6-Luna or Haiku 5.5 (they may also be serving lower than advertised rates under launch day demand) - BUT at its price it ends up being near the Pareto front, a few points behind Haiku 5.5 in performance at similar cost and speed, and closer to GPT-6-Luna in performance (while GPT-6-Luna is still cheaper in practice) - The main issue is polish -- we recorded more errors than similarly priced autoregressive transformers for things like syntax and reaching output token limits without an answer. Overall though, it's a dramatic improvement from the recent Mercury 2.5 model in a VERY short time, going from ~1 year behind the frontier to closely trailing GPT-6-Luna in a little over a month. They're serving at a price and speed that competes with the big boys without the same economies of scale. It absolutely proves the scalability of diffusion. BULLISH.