Muse Glimmer optimization reports on DGX Spark

# Muse-Glimmer-30B Benchmark Reports

Benchmark and optimization reports for Muse-Glimmer-30B and comparison models on NVIDIA DGX Spark (GB10).

## Reports

1. **[Glimmer Optimization Study](01-glimmer-optimization.md)** — Full optimization journey: BF16 → Q4 kquant → DFlash speculative decoding (10× speedup), plus NVFP4 via SGLang. Includes 8-config autoresearch sweeps for both frameworks.

2. **[Glimmer vs Qwen Comparison](02-glimmer-vs-qwen.md)** — Head-to-head comparison with Qwen3.6-27B: decode speed, GSM8K, HumanEval pass@1, architecture impact, and use-case recommendations.

3. **[Nemotron Three-Way Comparison](03-nemotron-three-way-comparison.md)** — Nemotron-3.5-Lightning-30B-A3B (MoE, NVFP4, vLLM) vs Glimmer and Qwen. Nemotron dominates: 95.6 tok/s (2× faster), 100% GSM8K, 92% HumanEval. Includes MoE architecture analysis and DFlash vs MTP comparison.

## Key Results

| Config | Framework | Decode (tok/s) | GSM8K | HumanEval |
|--------|-----------|:--------------:|:-----:|:---------:|
| **Nemotron NVFP4** | **vLLM** | **95.6** | **100%** | **92%** |
| Qwen Q4+DFlash | llama.cpp | 47.3 | 92% | 92% |
| Glimmer Q4+DFlash | llama.cpp | 38.3 | 96% | 82% |
| Glimmer NVFP4 | SGLang | 11.3 | 86% | 82% |

## Hardware
- NVIDIA DGX Spark (GB10, 128 GB unified memory)
- llama.cpp: CUDA 13.0, sm_121
- SGLang: FlashInfer SM120 backend, muse-glimmer branch (PR #34262)
- vLLM: 0.23.1rc1, FlashInfer 0.6.14, Marlin MoE backend

## Date
August 10–12, 2026