Qwen3.8-27B-heretic-dflash is a high-performance DFlash speculative drafter trained on multi-layer hidden representations from Qwen 3.8 27B Heretic models (including heretic-org/Qwen3.8-27B-heretic-ara and quantized variants like MXFP4).

By predicting future token blocks in parallel using single forward-pass target hidden representations, this drafter accelerates autoregressive generation on Apple Silicon (MLX) and NVIDIA GPUs by 2.6× to 2.8× with exact target model output fidelity.

🚀 Live Speculative Benchmarks (Apple Silicon MLX)

Evaluated against Qwen3.8-27B-heretic-ara-mxfp4 on real-world prompts with standard temperature:

Task / Domain	Prompt Scenario	Speculative Acceptance ($k$)	Acceptance Rate (%)	Decode Speed	End-to-End Speedup
Async Python Coding	Concurrent rate-limited web scraper with retry backoff & error handling	`3.01` tokens / round	`82.05%`	`21.9 tok/s`	`2.58×` faster (vs ~8.5 tok/s baseline)
Math & Reasoning	Recursive sequence formula derivation & proof ($\sum_{n=1}^{100} rac{1}{a_n}$)	`4.34` tokens / round	`80.50%`	`24.0 tok/s`	`2.82×` faster (vs ~8.5 tok/s baseline)

Draft Acceptance Rate: Over 81% of all proposed draft tokens are verified and accepted.
First-Token Acceptance Rate ($k \ge 1$): >95%, eliminating stalling during token generation.

📈 Multi-Round Continuous Training Progress

The model is iteratively trained on multi-domain instruction sequences (coding feedback, mathematical reasoning, system architecture, and multi-turn chat) using pure bfloat16 extracted hidden states:

Checkpoint Stage	Cumulative Sequences	Cumulative Tokens	Held-Out Val Loss	Block Accuracy ($k=15$)	Status
Untrained Base	`0`	`0`	`7.3247`	`7.7%`	Initial `z-lab` weights
Round 1 – 2	`6,000`	`12.4M`	`3.4830`	`19.3%`	Baseline Convergence
Round 3 – 5	`15,000`	`31.0M`	`3.8210`	`14.8%`	Cross-Domain Diversity
Round 6 – 8	`24,000`	`49.5M`	`3.6890`	`16.2%`	Reasoning & Proofs
Round 9 – 10	`30,000`	`61.9M`	`3.3995`	`17.8%`	Current Checkpoint (Live)
Round 11 – 17	`51,000`	`~105M`	Queued	Queued	Full 50k Step Schedule

🛠️ Architecture & Specifications

Drafter Architecture: DFlash Dual-Hidden Non-Autoregressive Drafter
Target Model: Qwen 3.8 27B Heretic / ARA
Hidden State Feature Dimension: $5 imes 5,120 = 25,600$ (concatenated target layers in pure bfloat16)
Context Length: 2,048 tokens + 16 block tokens
Block Size ($N$): 16 future tokens per speculative step
Storage Footprint: 3.46 GB safetensors

💻 How to Use with MLX / MLX-VLM

Launch mlx_vlm.server with --draft-model and --draft-kind dflash:

python3 -m mlx_vlm.server     --host 0.0.0.0     --port 8080     --model /path/to/Qwen3.8-27B-heretic-ara-mxfp4     --trust-remote-code     --kv-bits 4     --kv-quant-scheme uniform     --kv-group-size 64     --draft-model jfan/Qwen3.8-27B-heretic-dflash     --draft-kind dflash

Once running, send standard OpenAI-compatible /v1/chat/completions requests to enjoy accelerated 22–24+ tokens/second inference.