APEX widens MoE quants, adds I-Nano

APEX is starting to look like a real quantization strategy for MoE models, not just a one-off experiment on Qwen 3.5. The interesting part is less the model count than the repeated claim that routing-aware precision preserves long-context and code quality better than uniform quants.

• –The expansion across major MoE families suggests the approach is generalizing beyond one architecture
• –I-Nano is only plausible on MoEs because sparse expert activation can absorb extreme compression better than dense models
• –The reported memory savings are real but uneven; denser shared experts reduce the win, so architecture still matters
• –The strongest value proposition is practical deployment: fitting 30-70B-class MoEs onto a single consumer GPU without falling back to blunt uniform quantization
• –The evidence is promising, but still mostly self-reported and repo-owned benchmarks, so independent replication will matter