howtospark
models / laguna-m-1

226B params~23B activebfloat16MoE 256×16GQA 64:8262K ctxapache-2.0
Sparks
~469 GB
~356 GB
~243 GB
~130 GB
~82 GB
~81 GB
~76 GB
110 usable
REAP dialno pruning · 256 experts
Spec decode
Context33K / 262KKV
On the Sparks — 1× · 2-bit experts + NVFP4 dense · 33K ctx · FP16 KV~150K ctx fits usable
spark
34 GB free
Expert planes54* GB2-bit
110 usable
76* / 110 GB
69% of usable
Expert planesDense weightsKV cacheActivations + graphsOS reserve

* 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.

layer map
70 layersdense FFN×3MoE FFN×67
in
out

one MoE layer · 256 experts, 16 fire per token

each expert ≈ 13M params

resident: 67 MoE layers × 256 experts × 13M 215.8B

active/token: 67 × 16 × 13M 13.5B experts

+ 10.0B always-on dense (attention, embeddings, shared stack) = 23.5B active

Recipes (all recipes)

Worked deployments that run this model on DGX Spark — each shows how it fills each node's unified memory.

Optimize
Verdict

226B params, ~452 GB at bfloat16 and still ~113 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 256 routed experts and only 16 firing per token, REAP-style pruning cuts total params without touching active compute.

pruning to half the experts ≈ halves the routed-expert weights, ~62 GB at 4-bit

Split or stream it

Even 4-bit is ~113 GB against a 110 GB usable budget; shard across ≥2 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).

~64 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.

Add a speculative decoder

Decode is bandwidth-bound at ~23.3 tok/s ceiling; a trained EAGLE-style draft multiplies tokens per weight-read.

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.

Laguna-M.1 (poolside) — How to Spark · How To Spark