howtospark
models / Qwen/Qwen3.8-Flash-Next

Qwen/Qwen3.8-Flash-Next
180B params~62B activebfloat16MoE 512×10GQA 24:2262K ctxmultimodalother
Sparks
~371 GB
~281 GB
~191 GB
~101 GB
~63 GB
~101 GB
~71 GB
110 usable
REAP dialno pruning · 512 experts
Spec decode
Context33K / 262KKV
On the Sparks — 1× · 1-bit GGUF (dynamic) · 33K ctx · FP16 KV~511K ctx fits usable
spark
47 GB free
Dense weights17* GB2.3-bit
Expert planes34.7* GB2.3-bit
110 usable
63* / 110 GB
57% 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

+ 59.2B always-on dense (attention, embeddings, shared stack) = 61.6B active

Optimize
Verdict

180B params, ~360 GB at bfloat16 — doesn't fit natively, but ~90 GB at 4-bit does.

Make it fit

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

Quantize to 4-bit

At 360 GB native it doesn't fit the 110 GB budget; 4-bit weight-only quantization gets it to ~90 GB (params include vision/audio towers).

360 GB → ~90 GB

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 ~8.9 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.