wallawalla47/Qwen3.8-Flash-Next-DFlash2-Infernix

Qwen3.8-Flash-Next DFlash2 drafter for Infernix

AI disclaimer: this artifact, its conversion recipe, the engine that runs it and most of this card were made with AI (mostly Claude Opus 5.5). It is hobby work and is likely to be neither complete nor entirely accurate.

Inco AI's DFlash 2 drafter for Qwen3.8-Flash-Next, packed for Infernix, a C++/CUDA engine that runs Qwen3.8-Flash-Next on one RTX 5090. It is not a model by itself: Infernix loads it beside a Qwen3.8-Flash-Next artifact (--spec dflash2 --draft-model), and each decoding round it drafts a block of up to seven tokens in one pass, which the full model then verifies. Greedy output is the same as without it, and sampling keeps its distribution: only the speed changes.

One copy serves every Infernix Flash-Next artifact (FP8, NVIDIA NVFP4 Dense8, NVIDIA NVFP4, GSQ-RCO IQ3_S, Uncensored NVFP4): the drafter reads the main artifact's token embedding and its output or proposal head, so this file holds only the drafter's own layers.

Files

File	Bytes	What it is
Qwen3.8-Flash-Next-DFlash2-Infernix.infernix	776,075,776	the drafter (an Infernix companion artifact)
Qwen3.8-Flash-Next-DFlash2-Infernix.infernix.conversion.json	32,510	conversion report: sources, methods and formats per object
config.json		metadata only, so the Hub counts downloads; Infernix does not read it
LICENSE, NOTICE		the Apache License 2.0; attribution and changes

Get Infernix

Infernix is on GitHub: Wallawalla47/Infernix. Windows release zips are on its Releases page; on Linux (or to build the latest source on Windows) follow the build section of its README. DFlash2 for Qwen3.8-Flash-Next needs a build whose infernix-serve --help lists --draft-model; older builds refuse this file.

You also need a Qwen3.8-Flash-Next artifact for Infernix and an n-gram volume: the model cards linked above list them, and the Flash-Next guide covers everything else (KV formats, the prefix cache, the SSD expert tier, images and video).

Use

Download this file next to your main artifact:

hf download Wallawalla47/Qwen3.8-Flash-Next-DFlash2-Infernix Qwen3.8-Flash-Next-DFlash2-Infernix.infernix --local-dir <dir>

Then select DFlash2 instead of the built-in MTP drafter:

infernix-serve Qwen3.8-Flash-Next-NVIDIA-NVFP4-Dense8-Infernix-00001-of-00003.infernix \
  --ngram-volume <nvme>/Qwen3.8-Flash-Next-NVIDIA-NVFP4-Infernix.ngram \
  --max-context 262144 --kv-dtype int8 \
  --spec dflash2 --draft-model <dir>/Qwen3.8-Flash-Next-DFlash2-Infernix.infernix \
  --draft-tokens auto --lm-head-draft

• --spec dflash2 --draft-model FILE loads the drafter beside the main artifact (any Flash-Next artifact; Infernix checks that the drafter was made for this model).
• --draft-tokens auto (recommended): every round drafts the block of 7 the drafter was trained with (8 columns: the last accepted token and 7 drafts) and verifies only as many of them as the drafter's own probabilities, calibrated per request, make worthwhile. --draft-tokens N (1-7) verifies N every round, which was slower than MTP (see Performance); --spec-min-p P then verifies the drafts only up to the first whose probability is below P.
• --lm-head-draft scores the drafts with the main artifact's small proposal head (the 131,072 most frequent tokens, 4-bit) instead of its full output head; the figures below use it. Every Infernix Flash-Next artifact listed above has one.
• --ngram-draft-tokens 7 (copy proposals) works beside it, as with MTP.
• The same options work with the one-shot CLI (infernix) and infernix_bench.

DFlash2 replaces MTP: the main artifact's MTP drafter is not loaded (no --spec mtp), so the expert cache gets that VRAM back, less the drafter's 776 MB and a 40 MiB context ring per request lane. With --max-concurrency 2 or more, rounds of two or more requests decode without drafts (as MTP's do without the adaptive-speculation fast path); a request decoding alone speculates.

Quantization

Converted from Inco AI's checkpoint (revision a22a1744e85c17364dcc54a6a420af7135d4c7f6) with Infernix's recipe qwen3_8_flash_next_dflash2, stored as NInfer's official Qwen3.8-27B artifact stores its DFlash2 drafter:

Tensors	Format
Target-feature projection, attention query/key/value/output, MLP gate/up/down (46 projections; the context key/value projections share the query block's key/value words, as in the source)	q8_g32_fp16: 8-bit codes, one FP16 absmax scale per 32 weights
Dynamic-convolution kernel projections and base kernels, candidate selector projection and codebooks, every norm (45 tensors)	BF16, unchanged

A check against the checkpoint found every BF16 tensor byte-exact and every 8-bit weight within half a quantization step of its source value. The drafter only proposes tokens, so its precision can change how many drafts are accepted, never what is generated.

Performance

RTX 5090 on PCIe 5.0 x8, i9-13900K with 96 GB of RAM, INT8 KV, one request; 24 prompts of 1,280 tokens each continued for 384 tokens, from Infernix's own source code (code) and its Markdown documentation (docs), greedy or sampled at temperature 1.0 (top-k 20, top-p 0.95); both drafters with --draft-tokens auto and --lm-head-draft (2026-10-20). Decode tok/s, the mean of every run where there were several (Dense8 code: four DFlash2 runs at 209-217, eight MTP runs at 186-203):

Main artifact	Text	Plain	MTP (--spec mtp)	DFlash2 (this drafter)	DFlash2 vs MTP
NVIDIA NVFP4 Dense8	code, greedy	120	198	213	+8 %
NVIDIA NVFP4 Dense8	code, sampled		162	204	+26 %
NVIDIA NVFP4 Dense8	docs, greedy	121	164	179	+9 %
NVIDIA NVFP4 Dense8	docs, sampled	122	145	161	+11 %
GSQ-RCO IQ3_S	code, greedy	121	233	258	+11 %
GSQ-RCO IQ3_S	docs, greedy		192	210	+9 %
GSQ-RCO IQ3_S	docs, sampled		167	183	+10 %

With --draft-tokens auto each round drafts seven tokens and verifies only as many as pay for their verification: on this offloaded model a verified column costs about a third of a plain round in expert-cache misses, so verifying all seven every round (--draft-tokens 7) was slower than MTP. The drafter also leaves more VRAM to the expert cache than the MTP drafter: 9,070 cached experts against 8,740 on Dense8 (12,432 against 12,288 on IQ3_S, whose MTP experts are 2-bit).

How it works

DFlash 2 is a block-diffusion drafter: from the target's hidden features (the mean of the four hyper-connection streams of layers 5, 15, 25, 35 and 45), five sliding-window layers with dynamic convolutions turn a block of one known token and seven mask tokens into candidates for the next seven positions in one pass, and a candidate selector traces one path through the top 16 at each position. See the Inco AI blog post and Infernix's DFlash reference.

License

Apache License 2.0, as the source checkpoint (LICENSE); see NOTICE for the attribution and the changes made in this conversion. The main model it drafts for has its own license (the Qwen Community License 1.0 for Qwen3.8-Flash-Next).