You can deploy this model and start earning money today!

MirecX/
Qwen3.5-397B-A17B-L5H5-DFlash

Name: Qwen3.5-397B-A17B-L5H5-DFlash
Availability: InStock

safetensors

qwen3

custom_code

Views

No views yet

Qwen3.5-397B-A17B-L5H5-DFlash

Experimental DFlash speculative-decoding drafter for QuantTrio/Qwen3.5-397B-A17B-AWQ.

⚠️ Highly experimental — for tinkering, not production-grade. Recommend L3H5-DFlash over this one at the current training scale; L5H5 has more capacity but underperforms with only ~226 training shards.

Shape

5 drafter layers, hidden_size=4096, ~1.03B params, bf16
target taps: layers [2, 16, 30, 43, 57]
block_size=16
all full_attention

Eval (50 held-out shards, 200 blocks, greedy verification)

mean accept length: 1.20
k=1: 52.0%
cumulative top-1:

k 1 2 3 4 5 6 8 10 12 15
% 52.0 24.5 12.0 9.5 5.5 4.0 1.5 1.5 1.0 0.5

L5H5 needs significantly more training data (~1500+ shards expected) to outperform the smaller L3H5 sibling. Published for comparison and as a starting point for longer-schedule experiments.

Use with vLLM

--speculative-config '{"method":"dflash","model":"MirecX/Qwen3.5-397B-A17B-L5H5-DFlash","num_speculative_tokens":6}'

Training

Same corpus and recipe as L3H5 — 226 mixed shards, 10k steps, anchors=6, decay=0.85
best ckpt: step_10000
training: RTX 5080 16 GB, ~17 min

Capture path

Trained on hidden states from the patched dflash_extract plugin (full residual stream). See L3H5 README for the historical bug context.