howtospark
models / unsloth/Qwen3.8-Flash-Next-GGUF

177B params~59B activebfloat16MoE 512×10GQA 24:2262K ctxmultimodalotherGGUF
Sparks
~365 GB
~277 GB
~188 GB
~100 GB
~62 GB
~98 GB
~69 GB
110 usable
REAP dialno pruning · 512 experts
Spec decode
Context33K / 262KKV
On the Sparks — 1× · 1-bit GGUF (dynamic) · 33K ctx · FP16 KV~520K ctx fits usable
spark
48 GB free
Dense weights16.1* GB2.3-bit
Expert planes34.7* GB2.3-bit
110 usable
62* / 110 GB
56% 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
48 layerslinear attention×36full attention×12
in
out

one MoE layer · 512 experts, 10 fire per token · showing 256

each expert ≈ 5M params

resident: 48 MoE layers × 512 experts × 5M 120.8B

active/token: 48 × 10 × 5M 2.4B experts

+ 56.1B always-on dense (attention, embeddings, shared stack) = 58.5B active

Optimize
Quantized from

Qwen/Qwen3.8-Flash-Next

Analyze the original repo for the full architecture picture.

Verdict

177B params, ~354 GB at bfloat16 — doesn't fit natively, but ~88 GB at 4-bit does.

Make it fit

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

Pick the right GGUF quant

The repo already ships multiple GGUF quant levels — grab the largest that fits under 110 GB instead of quantizing yourself.

~88 GB target

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 ~9.3 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 ~26 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.