howtospark
models / Qwen/Qwen3.8-2.4T-A95B

2.4T params~122B activebfloat16MoE 512×10GQA 64:4262K ctxother
Sparks
~4913 GB
~3690 GB
~2467 GB
~1243 GB
~724 GB
~688 GB
~651 GB
110 usable
REAP dialno pruning · 512 experts
Spec decode
Context33K / 262KKV
On the Sparks — 1× · 2-bit experts + NVFP4 dense · 33K ctx · FP16 KV~0 ctx fits usable
spark
KV cache12.3* GBFP16
Dense weights37.7* GB4-bit
Expert planes592.7* GB2-bit
110 usable
+541 over
651* / 110 GB
592% 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
92 layerslinear attention×69full attention×23
in
out

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

each expert ≈ 50M params

resident: 92 MoE layers × 512 experts × 50M 2370.8B

active/token: 92 × 10 × 50M 46.3B experts

+ 75.4B always-on dense (attention, embeddings, shared stack) = 121.7B active

Optimize
Verdict

2446B params, ~4892 GB at bfloat16 and still ~1223 GB at 4-bit — one Spark can't hold it; you need surgery or more Sparks.

Make it fit

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

Prune experts (then quantize)

With 512 routed experts and only 10 firing per token, REAP-style pruning cuts total params without touching active compute.

pruning to half the experts ≈ halves the routed-expert weights, ~673 GB at 4-bit

Split or stream it

Even 4-bit is ~1223 GB against a 110 GB usable budget; shard across ≥12 Sparks or stream cold weights from NVMe.

Go below 4 bits

Codebook/outlier methods reach 2–3 bpw when 4-bit still overflows. — for MoE, the proven shape is a 2-bit sign-symmetric codebook on the routed experts with the dense stack kept at FP8 (vLLM-Moet).

~668 GB with 2-bit experts + FP8 dense (est.)

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 ~4.5 tok/s ceiling; a trained EAGLE-style draft multiplies tokens per weight-read.

Tame the KV cache

At 262144 tokens the KV cache alone is ~99 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.