Qwen3.8-27B-DFlash2 needs about 1.8 GB of memory at Q4_K_M (a 1.1 GB file) with an 8K context window, so it fits on a 4 GB GPU. This repository has 3 quants, from 1.1 GB (Q4_K_M) to 3.9 GB (BF16). Longer context needs more: at 32K tokens the KV cache alone adds about 0.7 GB.
Qwen3.8-27B-DFlash2: a dense, vision‑language draft model for speculative decoding
Qwen3.8-27B-DFlash2 is a DFlash 2 draft model built on Qwen3.8-27B. It runs inside a speculative decoding server and drafts blocks of tokens for the target model. The model supports vision‑language tasks and is compatible with llama.cpp with DFlash 2 support.
Good at
- Vision‑language understanding
- Speculative decoding drafting
- Long‑context reasoning
Limits
- No standalone language model; requires a target model
- Requires llama.cpp with DFlash 2 support
- Large VRAM for 8k context
Best for
- Integrating into speculative decoding pipelines
- Enhancing vision‑language inference
- Optimizing token drafting for large models
Who it’s for. Developers building inference engines that need efficient token drafting for vision‑language models. It is suitable for users who already have a target model and want to reduce decoding latency.
Quants & memory
| Quant | File | Est. memory | vs VRAM | Result | |
|---|---|---|---|---|---|
| Q4_K_M | 1.1 GB | 1.8 GB | Fits | Get file → | |
| Q8_0 | 2.1 GB | 2.7 GB | Fits | Get file → | |
| BF16 | 3.9 GB | 4.5 GB | Fits | Get file → |
For 8 GB VRAM, use Q4_K_M (1.14 GB, ~1.8 GB memory at 8k context). For 16 GB VRAM, Q8_0 (2.06 GB, ~2.7 GB memory). BF16 requires ~4.5 GB memory and is best for 32 GB VRAM setups. Memory estimates are approximate.
What GPU do you need for Qwen3.8-27B-DFlash2?
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 (2.1 GB file, about 2.7 GB in use).
- 6 GB VRAM: the best quant that fits is BF16 (3.9 GB file, about 4.5 GB in use).
Which Qwen3.8-27B-DFlash2 quant should you choose?
Q4_K_M is the usual starting point for Qwen3.8-27B-DFlash2: it balances file size and quality well. Here it is a 1.1 GB file.
The Q8_0 file is 2.1 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 Qwen3.8-27B-DFlash2
How much VRAM does Qwen3.8-27B-DFlash2 need?
About 1.8 GB at Q4_K_M with an 8K context window. That is the 1.1 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 Qwen3.8-27B-DFlash2 on an 8 GB GPU?
Yes, with BF16: the 3.9 GB file needs about 4.5 GB at 8K context. Larger quants in this repository do not fit in 8 GB.
What is the file size of Qwen3.8-27B-DFlash2 GGUF?
It depends on the quant: Q4_K_M is 1.1 GB, Q8_0 is 2.1 GB, BF16 is 3.9 GB.
What context length does Qwen3.8-27B-DFlash2 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 Qwen3.8-27B-DFlash2 GGUF?
The files are on Hugging Face in the z-lab/Qwen3.8-27B-DFlash2-GGUF repository. The download buttons on this page open the file on Hugging Face. ggufmodels.com does not host any model files.
What license does Qwen3.8-27B-DFlash2 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 Qwen3.8-27B-DFlash2 on 8 GB VRAM?
Yes, the Q4_K_M quant uses 1.14 GB file size and requires about 1.8 GB VRAM at 8k context.
Does this model support vision tasks?
Yes, it is a native vision‑language model that can understand images and videos.
What engine is needed to run this model?
llama.cpp compiled with DFlash 2 support (PR #27342) is required.