howtospark
models / nemotron-3-nano-30b-a3b

32B params~4.1B activebfloat16MoE 128×6GQA 32:2262K ctxother
Sparks
~73 GB
~57 GB
~41 GB
~26 GB
~19 GB
~18 GB
~18 GB
110 usable
REAP dialno pruning · 128 experts
Spec decode
Context33K / 262KKV
On the Sparks — 1× · 2-bit experts + NVFP4 dense · 33K ctx · FP16 KV~1.8M ctx fits usable
spark
92 GB free
110 usable
18* / 110 GB
16% 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
52 layersMoE FFN
in
out

one MoE layer · 128 experts, 6+1 fire per token

each expert ≈ 5M params

resident: 52 MoE layers × 128 experts × 5M 30.6B

active/token: 52 × 6 × 5M 3.1B experts

+ 947M always-on dense (attention, embeddings, shared stack) = 4.1B active

Eval scores (compare all)

Reported

78.3MMLU-Pro
73GPQA Diamond
68.3LiveCodeBench
33.3SciCode
10.6HLE
50MiniF2F
89.1AIME 2025
99.2AIME 2025 (tools)
75GPQA Diamond (tools)
15.5HLE (tools)
8.5Terminal-Bench Hard
38.8SWE-bench Verified
48τ²-Bench (airline)
56.9τ²-Bench (retail)
42.2τ²-Bench (telecom)
49τ²-Bench
53.8BFCL v4
71.5IFBench
38.5Multi-Challenge
67.7Arena-Hard-v2
35.9AA-LCR
92.9RULER @ 256k
91.3RULER @ 512k
86.3RULER @ 1M
59.5MMLU-ProX
86.2WMT24++ (en→xx)

Recipes (all recipes)

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

Optimize
Verdict

32B params, ~63 GB at bfloat16 — fits on one Spark as-is.

Make it fit

Get the weights (and headroom for KV) inside the 128 GB pool — the hard gate.

Lossless compression

BF16 weights entropy-code ~30% smaller bit-exact — smaller downloads and less bandwidth per token, no quality question at all.

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 ~33.6 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.

Tame the KV cache

At 262144 tokens the KV cache alone is ~14 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.

Nemotron 3 Nano 30B-A3B — How to Spark · How To Spark