While there has been a lot of discussion regarding the Nemotron Super family of models, I feel like the newest addition, the Nemotron Cascade 2 30B-A3B (which is *not* based on the Qwen architecture despite a similar size, it's a properly hybrid model based on Nemotron's own arch) has largely flown under the radar.
I've been running some evals on local models lately since I'm kind of tired of the "vibe feels" method of judging them. A combo that I quite like is HumanEval + ClassEval, simply because they're quick to run and complicated enough for most small models to still have noticeable differences. So, I gave mradermacher's IQ4_XS quant for a spin.
On HumanEval, Cascade 2 achieved a whopping 97.6%, leaving both medium Qwen3.5 models in the rear window. Similarly, it obtained a respectable 88% on ClassEval.
I'm going to run some more tests on this model, but I feel it deserves a bit more attention.
[link] [comments]
