Solar-Open2-250B-DSpark

A DSpark speculative-decoding draft model for Solar-Open2-250B-W4A8.

Model Detail

Field	Value
Method	DSpark (block-diffusion draft, block rejection sampling)
Target model	Solar-Open2-250B-W4A8
Block size / num_speculative_tokens	7 (updated from the earlier ns4 export)
Hidden layers	5
Attention geometry, RoPE, KV-head count, rms_norm_eps	inherited from target
Auxiliary hidden-state taps	layers 4, 12, 24, 36, 44 of the target — all full (softmax) attention layers; Solar-Open2-250B interleaves 12 full-attention layers among 36 linear-attention (KDA) layers, and only the full-attention layers are tapped
Training	on-policy (responses regenerated by the target itself), 3 epochs over a refreshed corpus (~1M examples, tool-call and multilingual oversampled); prompt mixture = public instruction/code/multilingual datasets + a private dataset (not published)

Benchmark accept length

Measured serving-time accept length (average accepted tokens per draft, 1 + accepted/drafts) on public benchmark question sets, via vLLM /metrics spec_decode_num_* counters:

Dataset	Questions	Accept length
GSM8K	1,319	4.75
HumanEval	164	4.63
MATH-500	500	4.18
MBPP (sanitized)	257	4.10
AIME 2026	30	3.47
MT-Bench	80	3.39
SWE-Rebench	50	2.92

Structured domains (math, code) accept more tokens per draft step than open-ended chat, consistent with the pattern reported for other DSpark draft models.

Usage

vllm serve vessl/Solar-Open2-250B-W4A8 \
  --served-model-name solar-open2-250b \
  --tensor-parallel-size 4 --enable-expert-parallel \
  --kv-cache-dtype fp8 \
  --enable-prefix-caching \
  --speculative-config '{"method":"dspark","model":"vessl/Solar-Open2-250B-DSpark","num_speculative_tokens":7}' \
  --reasoning-parser solar_open2 --tool-call-parser solar_open2 \
  --enable-auto-tool-choice \
  --port 8000

License

Distributed under the same Upstage Solar License as the Solar Open 2 base model and its W4A8 derivative, per the license's requirements for derivative AI models (name prefixed with "Solar," "Built with Solar" attribution).