models / jhu-clsp/mmBERT-small
~0.1B paramsfloat328K ctxmit
Sparks
Users
~9 GB
~8 GB
~8 GB
~8 GB
110 usable
Spec decode
Context8K / 8KKV
On the Sparks — 1× · 1-bit GGUF (dynamic) · 8K ctx · FP16 KV~3.0M ctx fits usable · ~368 users fit
spark
102 GB free
110 usable
8* / 110 GB
8% of usable
Dense 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
22 layerstransformer layer
in
out
all 22 layers fire per token — 140M active
Optimize
Verdict
0B params, ~1 GB at float32 — fits on one Spark as-is.
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 ~485.8 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.
Edit or steer it
One-shot weight math — no gradients, no corpus.