---
license: apache-2.0
inference: false
tags:
  - dflash2
  - draft-only
  - speculative-decoding
  - block-diffusion
  - draft-model
  - qwen3.8-27b-target
  - mlx
---

# DFlash2

This repository contains the promoted DFlash2 draft head used with
[`Qwen/Qwen3.8-27B`](https://huggingface.co/Qwen/Qwen3.8-27B). It is not a
fine-tune of the full Qwen target and must not be loaded as a standalone
language model. It has no tokenizer and no target weights: a compatible
DFlash2 runtime loads this head alongside the Qwen verifier.

The Hub base-model tree is intentionally omitted. A conventional
"Qwen finetune" label would misrepresent the artifact: only the small
speculative draft head is changed, while target-controlled output remains the
authority.

The head descends from the public Qwen3.8-27B DFlash2 draft model. It was
adapted with target-verified cycles from the incumbent legacy corpus and a
session-grouped Fable agentic corpus. The final checkpoint combines a
generation-2 Fable adaptation at strength `0.0625` with a generation-3
combined-domain adaptation at strength `0.1875`.

No trace dataset, prompt file, capture cycle, selector tensor, or W&B log is
included in this repository. Only the final draft checkpoint, configuration,
blend metadata, and evaluation summary are published.

## Acceptance results

All acceptance runs used greedy target-controlled generation on one Apple M4
Max through DFlash MLX. The target was the calibrated Qwen3.8-27B MXFP4
runtime. `Draft quantization` was disabled for this head.

| Paired gate | Incumbent | This head | Mean delta | Wins / ties / losses |
| --- | ---: | ---: | ---: | ---: |
| Legacy 24-prompt corpus | `0.6712239583` | `0.6718750000` | `+0.0006510417` | 5 / 14 / 5 |
| Fable 25-prompt held-out | `0.6471875000` | `0.6492187500` | `+0.0020312500` | 11 / 5 / 9 |

The Fable live-gate prompts were excluded from the generation-2 and combined
training splits. The legacy corpus had already been used to train the
incumbent, so it remained in the combined training capture for a paired
comparison.

## Exact parity

The exact 256-token baseline-versus-DFlash smoke gate passed:

- `token_match`: `true`
- smoke acceptance: `0.7109375`
- diagnostic baseline throughput: `21.9615` tok/s
- diagnostic DFlash throughput: `33.5111` tok/s

Throughput is diagnostic only. Thermal pressure and co-tenant machine load
were not controlled.

## Training summary

- combined capture: 11,273 target-verified cycles
- prompt groups: 127
- legacy contribution: 1,981 cycles from 24 prompts
- Fable contribution: 9,292 cycles from 103 training prompts
- excluded Fable live gate: 25 prompts
- teacher signal: target top-64 logits and greedy target tokens
- final full-run held-out prefix accuracy: `0.7932212528`
- final full-run held-out top-1 accuracy: `0.8714877986`

The capture teacher was the calibrated MXFP4 Qwen3.8-27B target runtime, not
the pinned BF16 Qwen3.8-27B checkpoint. This precision choice is part of the
training provenance.

## Checkpoint integrity

- `model.safetensors` SHA-256:
  `c90e0fc65dddbe4107b1b417d771c3be188c87197e89cf4b14415f56abc866f0`
- `config.json` SHA-256:
  `873e3556509b0da06e29654ba00d4944888d4b5e8a33afde25f7eb27d321e980`
- `blend.json` SHA-256:
  `e53a6bb5e890d7e3368053302189766c3ef68d2eb7887152ece1ea9eac4f64d5`

## Intended use and limits

This artifact is for lossless speculative decoding research with a compatible
DFlash2 implementation. Acceptance metrics do not establish broad task quality,
natural-EOS behavior, multilingual performance, long-context performance, or a
controlled thermal throughput win. The Fable held-out prompts come from the
same eval shard as the Fable training prompts, so that gate does not prove
out-of-domain generalization.