license: apache-2.0
library_name: transformers
pipeline_tag: text-generation
base_model: Dragoy/Swift-Qwen3.8-27B-abliterated-NVFP4
tags:
- dflash2
- speculative-decoding
- block-diffusion
- draft-model
- qwen3.8
- qwen3_5
- abliterated
- uncensored
- vllm
- sglang
- transformers
Swift-Qwen3.8-27B-abliterated-NVFP4-DFlash2
DFlash 2 draft model for
Dragoy/Swift-Qwen3.8-27B-abliterated-NVFP4
— the standard HF (NVFP4+FP8) checkpoint of the huihui-style abliterated
Swift-Qwen3.8-27b.
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 — 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. Decoding is
lossless: greedy output matches the target model exactly, and sampling
preserves the target distribution.
This repository is a byte-identical mirror of
z-lab/Qwen3.8-27B-DFlash2 @50307d4c — the same revision that was baked into thedflash2 component of the
NInfer artifact, so both
stacks (standard HF/vLLM/SGLang and NInfer) use one drafter.
Specs
| Target | Dragoy/Swift-Qwen3.8-27B-abliterated-NVFP4 |
| Source | z-lab/Qwen3.8-27B-DFlash2 @ 50307d4c (Apache-2.0) |
| Draft | 5-layer Qwen3-style transformer, 1.92B params, BF16, sliding-window attention (2048) |
| DFlash 2 config | block size 8 (7 draft tokens per verification step), mask token 248070, target layers [5, 19, 33, 47, 61] |
model.safetensors |
3.849 GB — sha256 67fc76d68dc5a9415511a4f394ef744d67510cd20e93b37cc2cc7d28e4bab65c |
config.json |
sha256 873e3556509b0da06e29654ba00d4944888d4b5e8a33afde25f7eb27d321e980 |
| Published | 2026-09-18 |
Quick Start
Serve the target with vLLM:
vllm serve Dragoy/Swift-Qwen3.8-27B-abliterated-NVFP4 \
--max-model-len 32768 --gpu-memory-utilization 0.9 \
--speculative-config '{{
"method": "dflash",
"model": "Dragoy/Swift-Qwen3.8-27B-abliterated-NVFP4-DFlash2",
"num_speculative_tokens": 7
}}'
Or with SGLang:
python -m sglang.launch_server \
--model-path Dragoy/Swift-Qwen3.8-27B-abliterated-NVFP4 \
--speculative-algorithm DFLASH \
--speculative-draft-model-path Dragoy/Swift-Qwen3.8-27B-abliterated-NVFP4-DFlash2 \
--speculative-num-draft-tokens 8
Verified on target
SMOKE_RESULTS_PLACEHOLDER
Notes
- The drafter was trained on the base
Qwen/Qwen3.8-27B; the target is
an abliterated derivative (refusal-direction removal, format/weight change,
no fine-tuning), so acceptance is expected to be close to the
base-model numbers. - DFlash 2: blog ·
github.com/z-lab/dflash
License
Draft weights: Apache-2.0
(z-lab/Qwen3.8-27B-DFlash2).
The target model is under the Swift Open License v1.0 — see the
target repository for the full chain.