Satgoy152/Muse-Glimmer-30B-DSpark-Coding

Muse Glimmer 30B DSpark Coding

A coding-specialized DSpark drafter for Muse Glimmer 30B, fine-tuned from DaoCloud/Muse-Glimmer-30B-DSpark on long-horizon, on-policy software-engineering traces.

This is a speculative-decoding drafter, not a standalone language model. Pair it with Muse Glimmer 30B as the verifier. The default checkpoint at the repository root is the 49K-context run; the 32K-context comparison checkpoint is retained under checkpoints/32k-context.

Results

Evaluation uses Muse-Glimmer-Terminal-Bench-Eval, the same BF16 Muse Glimmer 30B verifier, greedy decoding, and 15 speculative tokens. The table below reports the valid per-request mean acceptance length and pooled decode throughput at concurrency 1. Higher is better.

Acceptance and throughput

Drafter	Acceptance length (per request)	Change vs community DSpark	Pooled tok/s	Change vs community DSpark	Speedup vs no speculation
DFlash (official)	4.436	+10.3%	190.9	+3.8%	3.06×
DFlash2	6.306	+56.8%	249.6	+35.7%	4.00×
DSpark (community)	4.022	reference	183.9	reference	2.95×
This model — 32K context	6.938	+72.5%	239.7	+30.3%	3.84×
This model — 49K context	6.919	+72.0%	238.6	+29.7%	3.82×

Acceptance length is calculated per request as:

acceptance_length = 1 + accepted_draft_tokens / speculative_steps

The values above are then averaged over requests. Step-weighted acceptance is intentionally not reported.

Speedup across concurrency

Speedup is pooled decode throughput relative to the no-speculation control at the same concurrency.

Drafter	c1	c2	c8	c32
DFlash (official)	4.04×	3.69×	3.34×	1.05×
DFlash2	4.00×	3.93×	3.39×	1.79×
DSpark (community)	2.95×	2.82×	2.48×	1.31×
This model — 32K context	3.84×	3.62×	3.21×	1.77×
This model — 49K context	3.83×	3.55×	3.21×	1.79×

These benchmarks measure decoding efficiency, not task correctness. Throughput depends on hardware, runtime revisions, prompt/output lengths, and serving configuration.

Training recipe

Setting	Value
Verifier	meta-models/Muse-Glimmer-30B (BF16)
Initialization	DaoCloud/Muse-Glimmer-30B-DSpark
Training data	Satgoy152/Muse-Glimmer-SWE-Gym-2k
Evaluation data	Satgoy152/Muse-Glimmer-Terminal-Bench-Eval
Hidden-state layers	[2, 14, 26, 38, 50]
DSpark heads	Markov rank 256 + confidence head
Proposals	block_size=15, sample_from_anchor=true
Packed sequence length	49,152 tokens (default); 32,768 tokens (comparison checkpoint)
Loss	0.1 CE + 0.9 TV; confidence BCE weight 1.0
Optimizer	Muon for eligible 2D matrices; AdamW for the remaining parameters
Learning rate	Muon 1e-4; AdamW 5e-5
Schedule	Cosine decay, 3% warmup
Training	1 epoch

Checkpoints

Checkpoint	Location	Size	SHA-256
49K context (default)	Repository root	5,318,887,946 bytes	3e872ce337215113ed259802a4f0339bcd70af447da02039364d600a08aa62f2
32K context	checkpoints/32k-context	5,318,887,946 bytes	a71167bb97472bcb812feef9d8cc541cec66ad1df9a8821d44f7efd989aee84a

Example: serve with vLLM

vllm serve meta-models/Muse-Glimmer-30B \
  --tensor-parallel-size 1 \
  --enable-auto-tool-choice \
  --tool-call-parser muse_glimmer \
  --reasoning-parser muse_glimmer \
  --speculative-config \
  '{"method":"dspark","model":"Satgoy152/Muse-Glimmer-30B-DSpark-Coding","num_speculative_tokens":15}'

Notes

• The model was trained for coding and agentic workloads. General-chat, multilingual, and multimodal speculative decoding have not been comprehensively evaluated.
• The checkpoint was trained and evaluated with development versions of vLLM and Speculators. Runtime support may vary by release.
• The confidence head is included in the checkpoint, but the evaluated runtime does not use it for confidence-scheduled verification.
• Reproduction, training, evaluation, and serving code: Satgoy152/muse-glimmer-dspark.

Released under Apache-2.0, matching Muse Glimmer and the warm-start checkpoint.