This checkpoint adds a rank-256 Markov head and a confidence head to the warm-started DFlash backbone — about 103M new parameters, growing the checkpoint from 4.76 GiB to 4.95 GiB (+4%) — then trains on 99,984 Muse-Glimmer-generated, on-policy OpenPerfectBlend conversations. Its main benefit is much stronger acceptance deeper into a 15-token proposal.
Results
The following results use the same BF16 Muse Glimmer 30B verifier, greedy decoding, and 15 speculative proposals per draft call. Higher accepted length is better.
Speculators evaluation subsets
Subset
Official DFlash
DSpark
Change
HumanEval
3.887
4.641
+19%
Math reasoning
3.867
5.237
+35%
QA
3.192
3.768
+18%
Question
2.963
3.441
+16%
RAG
4.048
5.182
+28%
Summarization
3.551
4.568
+29%
Tool call
3.087
3.805
+23%
Translation
3.627
4.948
+36%
Writing
2.974
3.455
+16%
Open-source benchmarks
Subset
Official DFlash
DSpark
Change
GSM8K
3.734
4.914
+32%
MATH500
4.139
4.869
+18%
MBPP
3.647
4.390
+20%
AIME 2026
3.870
4.271
+10%
Accepted length is calculated from the raw server counters:
Per-position acceptance rate on all 13 subsets: Official DFlash vs DSpark
Per-position acceptance curves for the same runs as the tables above. DSpark shows slightly lower position-0 acceptance but substantially stronger acceptance deeper into the proposal, with the largest gains toward the tail.
Warm start
The inherited five-layer DFlash body is already trained, while the Markov and confidence heads are newly initialized. Applying the same 6e-4 peak learning rate to every parameter caused the warm-started body to lose some early-token accuracy during the high-LR phase. The released run uses:
text
1DFlash body: 3e-4
2Markov/confidence heads: 6e-4
Training curves comparing uniform and differential learning rates
Training recipe
Setting
Value
Verifier
Muse Glimmer 30B BF16
Initialization
Official Muse Glimmer five-layer DFlash assistant
Draft
5×Qwen3 GQA, causal SWA 2048, BF16
Hidden-state IDs
[2, 14, 26, 38, 50]
DSpark heads
Markov rank 256 + confidence head
Proposals
block_size=15, sample_from_anchor=true
Data
99,984 conversations / 148,900 train-turn rows
Reasoning mix
low 64,997; medium 10,000; high 19,991; xhigh 4,996
This technical preview was trained and validated with patched development versions of vllm-project/speculators and vllm-project/vllm. Stock releases may not yet include the required Muse Glimmer / DFlash / DSpark support.
The model was trained with a maximum proposal length of 15 tokens, but num_speculative_tokens is a serving-time tuning parameter. The optimal value should be determined based on workload, hardware, and latency.
The confidence head is included in the checkpoint, but the tested vLLM runtime does not yet use it for confidence-scheduled verification.
Multilingual, agentic, and multimodal speculative decoding has not yet been benchmarked and is left for future work.
Released under Apache-2.0, matching Muse Glimmer and its official DFlash assistant.
