---

base_model: incoai/Muse-Glimmer-30B-DFlash2

tags:

• transformers
• safetensors
• qwen3
• dflash2
• speculative-decoding
• block-diffusion
• draft-model
• sglang
• vllm
• text-generation
• base_model:meta-models/Muse-Glimmer-30B
• base_model:finetune:meta-models/Muse-Glimmer-30B
• license:apache-2.0
• text-generation-inference
• region:us

---

Muse Glimmer 30B DFlash2 GGUF

GGUF quantizations of Inco AI DFlash2 draft model for Muse Glimmer 30B.

Use with BeeLlama.cpp, a llama.cpp fork with advanced quantization features.

---

Muse-Glimmer-30B-DFlash2

Blog | GitHub

This repository contains the DFlash 2 draft model for

meta-models/Muse-Glimmer-30B.

It is not a standalone language model: it runs inside a speculative

decoding server and drafts tokens for the target model to verify. It is

finetuned from

meta-models/Muse-Glimmer-30B-assistant,

the official DFlash drafter Meta ships with the model. The checkpoint is also

mirrored at z-lab/Muse-Glimmer-30B-DFlash2.

DFlash 2 is a block-diffusion drafter for speculative decoding. It predicts

a whole block of tokens in a single pass and keeps the top candidates at

every position. A lightweight selector then traces one coherent path through them.

Two-tap dynamic convolutions in the backbone keep the draft from decaying

toward the end of the block. Decoding is lossless: greedy output

matches the target model exactly, and sampling preserves its distribution.

<div align="center">

<img src="assets/dflash2-figure.png" alt="DFlash 2: parallel block drafting with a candidate path selector" width="100%">

</div>

Quick Start

Serve with SGLang:

pip install "sglang[all] @ git+https://github.com/sgl-project/sglang.git#subdirectory=python"

python -m sglang.launch_server \
  --model-path meta-models/Muse-Glimmer-30B \
  --speculative-algorithm DFLASH \
  --speculative-draft-model-path incoai/Muse-Glimmer-30B-DFlash2 \
  --speculative-num-draft-tokens 16

Or with vLLM:

pip install -U "vllm @ git+https://github.com/vllm-project/vllm.git@refs/pull/52816/head"

vllm serve meta-models/Muse-Glimmer-30B \
  --speculative-config '{
    "method": "dflash",
    "model": "incoai/Muse-Glimmer-30B-DFlash2",
    "num_speculative_tokens": 15
  }'

See the blog post for other engines and more details.

Evaluation

• Runtime: SGLang on one NVIDIA H200, with FlashAttention 3 for target and draft attention
• Speculation block size: 16 (15 draft tokens per verification step)
• Sampling: Muse's officially recommended parameters (temperature 1.0, top-p 0.95, top-k 64), with high reasoning strength
• Maximum new tokens: 4096
• Prompts: benchmark formatting from z-lab/dflash

We compare autoregressive decoding, the official DFlash drafter

(meta-models/Muse-Glimmer-30B-assistant),

a community DSpark drafter

(DaoCloud/Muse-Glimmer-30B-DSpark),

and DFlash 2. All speculative methods propose fifteen draft tokens per

verification step.

Acceptance Length

Acceptance length is the per-request mean of completion tokens divided by verification steps.

Higher is better.

| Task | Official DFlash | DSpark | DFlash 2 |

| :--- | ---: | ---: | ---: |

| GSM8K | 5.43 | 5.45 | 6.57 |

| MATH-500 | 5.39 | 5.01 | 6.56 |

| HumanEval | 4.11 | 4.33 | 5.66 |

| MBPP | 3.74 | 4.02 | 5.30 |

| MT-Bench | 3.52 | 3.59 | 4.42 |

Throughput

Throughput is total output tokens divided by end-to-end wall time.

Each cell shows output tok/s (speedup vs. autoregressive).

Concurrency 1

| Task | Autoregressive | Official DFlash | DSpark | DFlash 2 |

| :--- | ---: | ---: | ---: | ---: |

| GSM8K | 63.9 | 247.8 (3.88×) | 236.5 (3.70×) | 293.7 (4.59×) |

| MATH-500 | 64.0 | 246.3 (3.85×) | 218.4 (3.41×) | 295.5 (4.62×) |

| HumanEval | 65.1 | 210.5 (3.23×) | 201.4 (3.09×) | 266.2 (4.09×) |

| MBPP | 63.9 | 196.8 (3.08×) | 192.7 (3.02×) | 264.8 (4.14×) |

| MT-Bench | 64.0 | 164.6 (2.57×) | 159.7 (2.49×) | 197.4 (3.08×) |

Concurrency 8

| Task | Autoregressive | Official DFlash | DSpark | DFlash 2 |

| :--- | ---: | ---: | ---: | ---: |

| GSM8K | 476.6 | 1,574.4 (3.30×) | 1,456.1 (3.06×) | 1,816.6 (3.81×) |

| MATH-500 | 466.0 | 1,582.9 (3.40×) | 1,386.0 (2.97×) | 1,859.3 (3.99×) |

| HumanEval | 499.9 | 1,419.8 (2.84×) | 1,315.4 (2.63×) | 1,784.9 (3.57×) |

| MBPP | 491.6 | 1,278.9 (2.60×) | 1,266.6 (2.58×) | 1,719.7 (3.50×) |

| MT-Bench | 470.0 | 1,078.4 (2.29×) | 1,052.6 (2.24×) | 1,288.9 (2.74×) |

Concurrency 32

| Task | Autoregressive | Official DFlash | DSpark | DFlash 2 |

| :--- | ---: | ---: | ---: | ---: |

| GSM8K | 1,705.6 | 2,330.3 (1.37×) | 2,301.7 (1.35×) | 2,818.3 (1.65×) |

| MATH-500 | 1,710.2 | 2,427.3 (1.42×) | 2,185.0 (1.28×) | 2,869.6 (1.68×) |

| HumanEval | 1,798.1 | 2,170.0 (1.21×) | 2,068.9 (1.15×) | 2,780.2 (1.55×) |

| MBPP | 1,717.5 | 1,964.6 (1.14×) | 2,006.5 (1.17×) | 2,685.4 (1.56×) |

| MT-Bench | 1,721.8 | 1,668.0 (0.97×) | 1,627.2 (0.95×) | 1,975.5 (1.15×) |

Citation

If you find DFlash 2 useful, please cite:

@misc{inco2026dflash2,
  title  = {{DFlash 2: Keep Drafting Parallel}},
  author = {{Inco AI}},
  year   = {2026},
  month  = {August},
  url    = {https://inco.ai/blog/dflash2/}
}

Please also cite the original DFlash paper:

@inproceedings{chen2026dflash,
  title     = {{DFlash: Block Diffusion for Flash Speculative Decoding}},
  author    = {Chen, Jian and Liang, Yesheng and Liu, Zhijian},
  booktitle = {International Conference on Machine Learning (ICML)},
  year      = {2026}
}