SupersonicLabs/Julia-1 — How to Spark

models / SupersonicLabs/Julia-1

0.1B paramsapache-2.0

Sparks

~9 GB

~8 GB

110 usable

Spec decode

On the Sparks — 1× · 1-bit GGUF (dynamic)

spark

102 GB free

110 usable

8* / 110 GB

7% of usable

Dense weightsActivations + 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.

Optimize

Quantized from

Analyze the original repo for the full architecture picture.

Verdict

0B params, ~1 GB at native precision — fits on one Spark as-is.

Make it fast

Once it fits: fewer bytes per token and fewer decode steps against 273 GB/s.

Decode is bandwidth-bound at ~473 tok/s ceiling; a trained EAGLE-style draft multiplies tokens per weight-read.

Change what the model does — adapt, edit, or steer it — independent of size.

LoRA/QLoRA adapts behavior on one Spark without touching the base weights.

One-shot weight math — no gradients, no corpus.