MAMBAs in the Wild?

I posted this as a reply on BlueSky sometime recently, but I’m a touch baffled by the lack of consumer facing stats space/diffusion/hybrid models. NVIDIA pops out a new Nemotron (hybrid MAMBA Transformer) model every so often, but I usually see limited discourse online. I could be following the wrong people—reading the wrong articles—but I really think these releases would be more pervasive. The throughput gains seem like a major QOL improvement—at least on paper. I’ve tried Mercury Lab’s diffusion model a while ago and it was pretty fast.

Honestly I feel the same thing about Cerebras/Groq (that they should be bigger/more common), though I’m more hopeful they enjoy larger adoption as time passes, as opposed to new architectures (in the short term). I’m sure there are plenty of technical difficulties which I’m missing (notwithstanding the difficulty of moving away from CUDA which I think is probably necessitated by swapping to Cerebras’s ASICs?) but I’d love to see 1k tok/s interfaces hit the consumer space and force providers to match. Though I don’t have any money on it (yet!) I will maintain that the transformer will be ditched sometime, maybe not soon given the amount of research effort being pumped into it, but there will be a change. If I could tell you when, I’d make a lot of money off it first.