Apple's Stepped MoE: one model scales 1-4B parameters, beating dense counterparts
apple · hf · 2026-10-08
Apple published Stepped MoE on Hugging Face, a unified framework combining elastic architectures with sparsely gated routing so a single model adapts to both deployment constraints and task requirements.
- The backbone conditions on context and target efficiency specs, activating task-relevant parameters within elastically nested sub-networks for fine-grained accuracy-efficiency control at inference.
- One model can run at 1/2/3/4B parameter levels, scoring 2-5% higher than dense counterparts on knowledge-intensive benchmarks while matching static versions and delivering similar latency.
- Parameter sharing saves on-device disk space, enabling flexible serving based on available DRAM and compute — aimed at edge deployment.
More from Infra
- exe.dev explains crossing the hyper-thread boundary: core scheduling cookies for VM isolation — davidcrawshaw · 2026-10-08
- GitHub goes down; engineers confirm a fix is in progress — iannuttall · 2026-10-08
- Codex Cloud shipped with day-0 Tailscale support — and Tailscale didn't even know — pvncher · 2026-10-08
- Analyst raises Marvell 2030 revenue estimate to $56B after investor day on AI connectivity upside — BenBajarin · 2026-10-08
- NVIDIA's Vera Rubin roadmap signals the AI race is becoming an infrastructure race — mikeflache · 2026-10-08
- Cursor power user burns 1.5T tokens in 30 days — $132M/year at sticker price — zeeg · 2026-10-08