DFlash needs about 1.7 GB of memory at Q4_K_M (a 1.0 GB file) with an 8K context window, so it fits on a 4 GB GPU. This repository has 5 quants, from 1.0 GB (Q4_K_M) to 3.5 GB (BF16). Longer context needs more: at 32K tokens the KV cache alone adds about 0.7 GB.
DFlash: a lightweight drafting component for Qwen3.6-27B
DFlash is a draft model that uses block diffusion for efficient speculative decoding. It is designed to work with the target model Qwen/Qwen3.6-27B. The model is still under training and inference support may be incomplete.
Good at
- Efficient parallel drafting
- Speculative decoding
- Low memory footprint
Limits
- Requires Qwen3.6-27B as target
- Inference engine support may be limited
- No vision or multimodal support
Best for
- Speed‑critical drafting tasks
- Low‑latency inference setups
- Research on speculative decoding
Who it’s for. Developers building inference pipelines that need a fast drafting component for Qwen3.6-27B. Researchers exploring block diffusion and speculative decoding techniques.
Quants & memory
| Quant | File | Est. memory | vs VRAM | Result | |
|---|---|---|---|---|---|
| Q4_K_M | 1.0 GB | 1.7 GB | Fits | Get file → | |
| Q5_K | 1.2 GB | 1.9 GB | Fits | Get file → | |
| Q6_K | 1.4 GB | 2.1 GB | Fits | Get file → | |
| Q8_0 | 1.8 GB | 2.5 GB | Fits | Get file → | |
| BF16 | 3.5 GB | 4.1 GB | Fits | Get file → |
For 8 GB VRAM, use Q4_K_M (1.03 GB, ~1.7 GB RAM). For 12 GB VRAM, Q5_K (1.23 GB, ~1.9 GB RAM) is suitable. If more memory is available, Q6_K or Q8_0 provide higher precision.
What GPU do you need for DFlash?
These figures add the file size, an fp16 KV cache at an 8K context window (worked out from this model’s own header) and about 0.5 GB for the runtime. They are estimates, so leave a little headroom, and expect to use less context or a smaller quant if you are right at the limit.
- 4 GB VRAM: the best quant that fits is Q8_0 (1.8 GB file, about 2.5 GB in use).
- 6 GB VRAM: the best quant that fits is BF16 (3.5 GB file, about 4.1 GB in use).
Which DFlash quant should you choose?
Q4_K_M is the usual starting point for DFlash: it balances file size and quality well. Here it is a 1.0 GB file.
The Q8_0 file is 1.8 GB, about 1.8 times the size of Q4_K_M. It keeps the weights closer to the original, which only matters if you have the memory to spare.
New to quant names? Read GGUF quantization explained.
Specifications
Architecture from the GGUF header
Source
Frequently asked questions about DFlash
How much VRAM does DFlash need?
About 1.7 GB at Q4_K_M with an 8K context window. That is the 1.0 GB file, plus the KV cache and roughly 0.5 GB of runtime overhead. Smaller quants need less and longer context needs more.
Can I run DFlash on an 8 GB GPU?
Yes, with BF16: the 3.5 GB file needs about 4.1 GB at 8K context. Larger quants in this repository do not fit in 8 GB.
What is the file size of DFlash GGUF?
It depends on the quant: Q4_K_M is 1.0 GB, Q5_K is 1.2 GB, Q6_K is 1.4 GB, Q8_0 is 1.8 GB, BF16 is 3.5 GB.
What context length does DFlash support?
The GGUF header lists a trained context length of 262,144 tokens (256K). The KV cache grows with the context you use: about 0.2 GB at 8K and 0.7 GB at 32K.
Where can I download DFlash GGUF?
The files are on Hugging Face in the Alittlehammmer/Qwen3.6-27B-DFlash-GGUF-llama.cpp repository. The download buttons on this page open the file on Hugging Face. ggufmodels.com does not host any model files.
What license does DFlash use?
Hugging Face lists the license as apache-2.0. Licenses can differ between the quantized files and the original model, so check the repository and the base model before commercial use.
Can I run DFlash on 8 GB VRAM?
Yes, the Q4_K_M quant uses about 1.7 GB of RAM at 8 k context, fitting within 8 GB VRAM.
Does DFlash support vision tasks?
No, DFlash is a text‑only drafting model and does not include vision capabilities.
Do I need a special inference engine?
It must be used with Qwen3.6-27B and may require modified vLLM or SGLang setups due to SWA layers.