howtospark
models / solar-open2-250b
Upstage logo

250B params~15B activebfloat16MoE 320×8GQA 64:81.0M ctxother
Sparks
~515 GB
~390 GB
~265 GB
~140 GB
~86 GB
~84 GB
~79 GB
110 usable
REAP dialno pruning · 320 experts
Spec decode
Context33K / 1.0MKV
On the Sparks — 1× · 2-bit experts + NVFP4 dense · 33K ctx · FP16 KV~189K ctx fits usable
spark
31 GB free
Expert planes60.4* GB2-bit
110 usable
79* / 110 GB
72% 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 layersMoE FFN
in
out

one MoE layer · 320 experts, 8+1 fire per token · showing 256

each expert ≈ 16M params

resident: 48 MoE layers × 320 experts × 16M 241.6B

active/token: 48 × 8 × 16M 6.0B experts

+ 8.7B always-on dense (attention, embeddings, shared stack) = 14.7B active

Eval scores (compare all)

Reported

86.2MMLU-Pro
86.3GPQA Diamond
28.8HLE
92.4LiveCodeBench
55.9ArtifactsBench
93.9HMMT 2026
95.7AIME 2026
61Multi-Challenge
80IFBench
62.3AA-LCR
70.4SWE-bench Verified
16.6APEX-Agents
58.2MCP-Atlas
19.6τ³ (banking)

Recipes (all recipes)

Worked deployments that run this model on DGX Spark — each shows how it fills each node's unified memory.

Optimize
Verdict

250B params, ~501 GB at bfloat16 and still ~125 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 320 routed experts and only 8 firing per token, REAP-style pruning cuts total params without touching active compute.

pruning to half the experts ≈ halves the routed-expert weights, ~69 GB at 4-bit

Split or stream it

Even 4-bit is ~125 GB against a 110 GB usable budget; shard across ≥2 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).

~69 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 ~37.1 tok/s ceiling; a trained EAGLE-style draft multiplies tokens per weight-read.

Tame the KV cache

At 1048576 tokens the KV cache alone is ~206 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.