license: apache-2.0 base_model:
- incoai/Qwen3.8-Flash-Next-DFlash2 base_model_relation: finetune pipeline_tag: text-generation inference: false tags:
- dflash2
- speculative-decoding
- draft-model
- gguf
- strata
- metal
DFlash2 drafter for DavidAU's Qwen3.8-Flash-Next Abliterated GSQ-RCO
A DFlash2 draft model fine-tuned for one target:
DavidAU/Qwen3.8-Flash-Next-Abliterated-Uncensored-Turbo-Brilliance-50x-Reasoning-Instruct-GSQ-RCO-GGUF
(IQ3_S). It starts from incoai/Qwen3.8-Flash-Next-DFlash2,
which was trained on the official Qwen3.8-Flash-Next. It is not a standalone language model: it drafts tokens that the
target verifies, so the output is the target's own (greedy output is identical with or without it).
Files
| File | What it is |
|---|---|
model.safetensors, config.json | The fine-tuned drafter, BF16, same layout as the incoai checkpoint. |
t3000-q8-dv.gguf | The drafter for Strata's --dflash on Mac: Q8_0, with the target's output head rows for the MTP draft vocabulary added (tools/dflash_head.py --draft-vocab). sha256 e5412a73583755eea8daa4c4c5793f5b827af1fbc0e76bcdd5cc181cdbf46eb1. |
Training
3000 steps, batch 32, AdamW with cosine decay and 100 warm-up steps, on the target's own features: the outputs of target layers 5/15/25/35/45 (mean of the four hyper-connection streams) and the target's continuations, captured from the IQ3_S model itself. Held-out drafts accepted per 7-token block: 3.568 for the incoai drafter, 3.796 after the fine-tune.
Use with Strata (Apple Silicon)
Needs the Metal engine with --dflash (opt-in; Strata PR #3457). In the config's args, in place of --mtp …:
--spec 4 --dflash t3000-q8-dv.gguf --spec-min-p 0.75
with STRATA_DFLASH_NATIVE=1 in env.
Measured
Apple M4 Max, 64 GB, macOS 26.5, expert arena 42 GB, greedy, against the MTP layer (spec 3, min-p 0.7) on the same machine. Two runs each for short prompts and a 29K-token prompt, one run for 102K.
| MTP | This drafter | |
|---|---|---|
| Short prompts, decode tok/s | 48.1, 46.9 | 49.7, 49.7 |
| 29K prompt, decode tok/s (cold) | 50.4, 50.0 | 51.0, 51.7 |
| 102K prompt, decode tok/s (cold / cached follow-up) | 46.8 / 44.4 | 50.5 / 45.1 |
The incoai drafter without this fine-tune was slower than MTP on this target. Not measured: sampling at temperature > 0, other Macs, other quantizations, CUDA or HIP.
License
Apache-2.0, like the incoai drafter and the target model.