Nemotron REAP cut hits AIME 90%+

This is a serious local-inference experiment, but the headline number should be read as a community benchmark until broader evals and reproduction land.

• –REAP pruning from 512 to 256 experts is the real story: it cuts deployment weight without giving up the sparse-MoE active-parameter profile
• –FP8 beats AWQ on quality but takes a roughly 40% throughput hit, making this a practical quality-vs-latency choice for math workloads
• –The included vLLM patch matters because expert routing edge cases still break real-world serving paths for unusual MoE shapes
• –Fine-tuning on about 270 AIMO3 and AstralMath problems means the AIME result is impressive, but narrow and potentially sensitive to prompt placement
• –Single-GPU 90%+ AIME-class math performance is exactly the kind of open-weights pressure that makes smaller, specialized reasoning models worth watching