* 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 · 256 experts, 10 fire per token
each expert ≈ 9M params
resident: 48 MoE layers × 256 experts × 9M ≈ 114.0B
active/token: 48 × 10 × 9M ≈ 4.5B experts
+ 3.5B always-on dense (attention, embeddings, shared stack) = 8.1B active
Eval scores (compare all)
Reported
Recipes (all recipes)
Worked deployments that run this model on DGX Spark — each shows how it fills each node's unified memory.
Laguna-S 2.1 FP8
poolside's agentic-coding Laguna-S 2.1 (118B MoE, ~8B active) in FP8 on two DGX Sparks (TP2), with a self-quantized W4A16 DFlash speculator — 60.9 tok/s single-stream (2.3x the no-draft baseline) at the full 262K context.
Laguna-S 2.1 NVFP4
poolside's agentic-coding Laguna-S 2.1 (118B MoE, ~8B active) in NVFP4 on one DGX Spark, with the matching NVFP4 DFlash speculator — 42.9 tok/s single-stream (2.3x the no-draft baseline) at the full 262K context.
Optimize
118B params, ~235 GB at bfloat16 — doesn't fit natively, but ~59 GB at 4-bit does.
Make it fit
Get the weights (and headroom for KV) inside the 128 GB pool — the hard gate.
Quantize to 4-bit
At 235 GB native it doesn't fit the 110 GB budget; 4-bit weight-only quantization gets it to ~59 GB.
235 GB → ~59 GB
Make it fast
Once it fits: fewer bytes per token and fewer decode steps against 273 GB/s.
Add a speculative decoder
Decode is bandwidth-bound at ~67.8 tok/s ceiling; a trained EAGLE-style draft multiplies tokens per weight-read.
Prune experts for speed
Fewer experts ⇒ smaller working set ⇒ better cache/page behavior, even when it already fits.
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.