Nemotron Cascade 2 posts strong code gains

This is the kind of release that can get buried under bigger model chatter, but it matters a lot for local inference: a compact open MoE with surprisingly sharp code results is exactly what people running on consumer hardware want to see.

• –NVIDIA’s official docs place the model on top of Nemotron-3-Nano’s hybrid Mamba-Transformer MoE lineage, so the “not just another Qwen fork” angle appears to hold.
• –The reported HumanEval and ClassEval numbers are impressive, especially coming from a quantized local run rather than a curated vendor benchmark.
• –The official model card also frames Cascade 2 as thinking/instruct-capable, with strong results across coding, math, tool use, and agentic evals, so this is broader than a code-only toy.
• –The big caveat is that this is still one community datapoint; it’s promising signal, not final proof, but it’s enough to justify a wider bake-off.