---

base_model: incoai/GLM-5.3-Flash-DFlash2

tags:

• transformers
• safetensors
• qwen3
• dflash
• dflash2
• speculative-decoding
• block-diffusion
• draft-model
• sglang
• text-generation
• base_model:zai-org/GLM-5.3-Flash
• base_model:finetune:zai-org/GLM-5.3-Flash
• license:cc-by-nc-nd-4.0
• text-generation-inference
• region:us

---

GLM 5.3 Flash DFlash2 GGUF

GGUF quantizations of Inco AI DFlash2 draft model for GLM 5.3 Flash.

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

---

GLM-5.3-Flash-DFlash2

Blog | GitHub

This repository contains the DFlash 2 draft model for

zai-org/GLM-5.3-Flash.

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

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

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.

Quick Start

Serve with SGLang:

pip install "sglang[all] @ git+https://github.com/sgl-project/sglang.git@refs/pull/36708/head#subdirectory=python"

sglang serve \
  --model-path zai-org/GLM-5.3-Flash \
  --trust-remote-code \
  --speculative-algorithm DFLASH \
  --speculative-draft-model-path incoai/GLM-5.3-Flash-DFlash2 \
  --speculative-draft-attention-backend fa4

See the blog post for more details.

Evaluation

• Runtime: SGLang on four NVIDIA GB300 GPUs (TP4), with TRT-LLM DSA and FlashInfer TRT-LLM MoE for the target, FP8 target KV cache, and FlashAttention 4 for DFlash 2 draft attention
• Speculation block size: 8 (7 draft tokens per verification step)
• Sampling: GLM-5.3-Flash's officially recommended parameters (temperature 1.0, top-p 0.95), with the default Max reasoning effort
• Maximum new tokens: 4096
• Samples: 128 at concurrency 1; 1,024 at concurrency 8 and 32

We compare autoregressive decoding, GLM-5.3-Flash's native MTP, and DFlash 2.

All speculative methods propose seven 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 | MTP | DFlash 2 |

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

| GSM8K | 5.06 | 5.78 |

| MATH-500 | 4.95 | 5.86 |

| HumanEval | 4.70 | 5.32 |

| MBPP | 4.26 | 4.85 |

| MT-Bench | 3.71 | 4.03 |

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 | MTP | DFlash 2 |

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

| GSM8K | 146.8 | 282.6 (1.93×) | 355.4 (2.42×) |

| MATH-500 | 157.5 | 323.2 (2.05×) | 438.9 (2.79×) |

| HumanEval | 166.6 | 323.5 (1.94×) | 436.8 (2.62×) |

| MBPP | 168.2 | 299.5 (1.78×) | 402.2 (2.39×) |

| MT-Bench | 169.3 | 231.1 (1.36×) | 293.2 (1.73×) |

Concurrency 8

| Task | Autoregressive | MTP | DFlash 2 |

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

| GSM8K | 582.6 | 825.5 (1.42×) | 922.9 (1.58×) |

| MATH-500 | 794.9 | 1,267.8 (1.59×) | 1,552.3 (1.95×) |

| HumanEval | 897.4 | 1,533.1 (1.71×) | 1,943.5 (2.17×) |

| MBPP | 895.8 | 1,428.0 (1.59×) | 1,821.8 (2.03×) |

| MT-Bench | 864.0 | 1,084.5 (1.26×) | 1,303.2 (1.51×) |

Concurrency 32

| Task | Autoregressive | MTP | DFlash 2 |

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

| GSM8K | 917.3 | 1,203.1 (1.31×) | 1,318.3 (1.44×) |

| MATH-500 | 1,635.6 | 2,362.7 (1.44×) | 2,717.0 (1.66×) |

| HumanEval | 2,089.8 | 3,511.4 (1.68×) | 4,198.4 (2.01×) |

| MBPP | 2,102.0 | 3,168.5 (1.51×) | 3,791.4 (1.80×) |

| MT-Bench | 1,831.5 | 2,317.0 (1.27×) | 2,675.0 (1.46×) |

License

This model is released under

CC BY-NC-ND 4.0

for research and evaluation. For commercial licensing, contact

contact@inco.ai.

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}
}