* 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 · 384 experts, 6+1 fire per token · showing 256
each expert ≈ 35M params
resident: 40 MoE layers × 384 experts × 35M ≈ 543.6B
active/token: 40 × 6 × 35M ≈ 8.5B experts
+ 219.6B always-on dense (attention, embeddings, shared stack) = 228.1B active
Optimize
763B params, ~763 GB at already fp8-quantized and still ~382 GB at 4-bit — one Spark can't hold it; you need surgery or more Sparks.
Make it fit
Get the weights (and headroom for KV) inside the 128 GB pool — the hard gate.
Prune experts (then quantize)
With 384 routed experts and only 6 firing per token, REAP-style pruning cuts total params without touching active compute (params include vision/audio towers).
pruning to half the experts ≈ halves the routed-expert weights, ~210 GB at 4-bit
Split or stream it
Even 4-bit is ~382 GB against a 110 GB usable budget; shard across ≥4 Sparks or stream cold weights from NVMe.
Go below 4 bits
Codebook/outlier methods reach 2–3 bpw when 4-bit still overflows. — for MoE, the proven shape is a 2-bit sign-symmetric codebook on the routed experts with the dense stack kept at FP8 (vLLM-Moet).
~356 GB with 2-bit experts + FP8 dense (est.)
Make it fast
Once it fits: fewer bytes per token and fewer decode steps against 273 GB/s.
Use its MTP heads
Ships 3 multi-token-prediction layer(s) — self-speculative decode with no separate draft model; check vLLM/SGLang support for this arch.
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.