Satgoy152/Muse-Glimmer-30B-DFlash2-Coding

Muse Glimmer 30B DFlash2 Coding

A coding-specialized DFlash2 drafter for Muse Glimmer 30B, fine-tuned from the converted z-lab/Muse-Glimmer-30B-DFlash2 checkpoint 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 final checkpoint is at the repository root; the mid-training checkpoint is retained under checkpoints/mid-step-1976.

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 DFlash2	Pooled tok/s	Change vs DFlash2	Speedup vs no speculation
DFlash (official)	4.436	−29.7%	190.9	−23.5%	3.06×
DSpark (community)	4.022	−36.2%	183.9	−26.3%	2.95×
DFlash2	6.306	reference	249.6	reference	4.00×
This model — mid (step 1,976)	7.266	+15.2%	259.7	+4.0%	4.16×
This model — final (step 3,956)	7.350	+16.6%	265.9	+6.5%	4.26×

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×
DSpark (community)	2.95×	2.82×	2.48×	1.31×
DFlash2	4.00×	3.93×	3.39×	1.79×
This model — mid (step 1,976)	4.17×	4.04×	3.43×	1.85×
This model — final (step 3,956)	4.26×	4.02×	3.57×	1.89×

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	Converted z-lab/Muse-Glimmer-30B-DFlash2 checkpoint
Training data	Satgoy152/Muse-Glimmer-SWE-Gym-2k
Evaluation data	Satgoy152/Muse-Glimmer-Terminal-Bench-Eval
Hidden-state layers	[2, 14, 26, 38, 50]
DFlash2 heads	Convolution kernel 2, group size 16, selector rank 256, top-k 16
Proposals	block_size=16, sample_from_anchor=false (15 speculative tokens)
Packed sequence length	32,768 tokens
Loss	0.1 CE + 0.9 TV; selector loss 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, 10% warmup
Training	1 epoch, 4 trainer ranks

Checkpoints

Checkpoint	Location	Size	SHA-256
Final, step 3,956 (default)	Repository root	5,544,328,456 bytes	f520710be26f8ffaaa4151622391e9302c8a4b1b72e50a0283abe3ac7c9431d0
Mid, step 1,976	checkpoints/mid-step-1976	5,544,328,456 bytes	09bd708bd15ca9d8326e5342185171e87dba7aef19dd910a535805bb658acbae

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":"dflash2","model":"Satgoy152/Muse-Glimmer-30B-DFlash2-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.
• Reproduction, training, evaluation, conversion, and serving code: Satgoy152/muse-glimmer-dspark.

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