* segment sizes marked with an asterisk are estimates pending a measured run
Estimated per-node footprint, weights + KV sharded across the Spark (tensor parallel). Overhead ~8 GB/node; KV at FP16.
one MoE layer · 192 experts, 8+1 fire per token
each expert ≈ 11M params
resident: 80 MoE layers × 192 experts × 11M ≈ 166.8B
active/token: 80 × 8 × 11M ≈ 12.1B experts
+ 5.2B always-on dense (attention, embeddings, shared stack) = 17.2B active
Recipes (all recipes)
Worked deployments that run this model on DGX Spark — each shows how it fills each node's unified memory.
Optimize
172B params, ~86 GB at already modelopt-quantized — fits on one Spark as-is.
Make it fast
Once it fits: fewer bytes per token and fewer decode steps against 273 GB/s.
Use its MTP heads
Ships 1 multi-token-prediction layer(s) — self-speculative decode with no separate draft model; check vLLM/SGLang support for this arch.
Prune experts for speed
Fewer experts ⇒ smaller working set ⇒ better cache/page behavior, even when it already fits.
Tame the KV cache
At 262144 tokens the KV cache alone is ~86 GB at FP16; quantize or evict.
Make it yours
Change what the model does — adapt, edit, or steer it — independent of size.
Fine-tune it
LoRA/QLoRA adapts behavior on one Spark without touching the base weights.
Edit or steer it
One-shot weight math — no gradients, no corpus.