This repository contains a DFlash2 draft model for ornith-ai/Ornith-1.5-35B-A3B. It is not a standalone language model: a compatible Ornith target checkpoint is required at inference time, and the target verifies every proposed token.

The checkpoint was trained on projected teacher features from 100,014,884 generated assistant completion tokens (108,729 conversations). The full generation dataset is released separately as jzinno/Ornith-1.5-35B-A3B-Nemotron-v2-100M.

Target responses were generated first. A separate offline teacher-forcing pass then loaded the revision-pinned BF16 Ornith teacher, ran each completed sequence once, and selected hidden layers 1, 6, 11, 16, 22, 27, 32, and 37. Feature capture used 8 NVIDIA H200 141GB HBM3e GPUs in independent contiguous partitions; the selected drafter was trained on 1 NVIDIA H200 141GB HBM3e GPU.

The inherited 2048-wide projection from z-lab/Qwen3.5-35B-A3B-DFlash was applied during that pass and kept frozen. Only the projected BF16 features, token IDs, and loss masks were cached. Drafter training therefore did not keep the full teacher resident; it loaded only Ornith's frozen embedding and language-model head with the warm-started draft network.

The Qwen3.5 checkpoint is a plain DFlash drafter. Its compatible parameters initialize the shared draft network before adaptation to the continued-pretrained Ornith target. The DFlash2-only grouped causal convolutions and path selector start as exact no-ops: each convolution has a unit self-tap with zero predecessor/dynamic correction, and the selector's successor codebook is zero so its scores reduce to the inherited DFlash logits. The upgraded network is therefore numerically equivalent to the plain DFlash warm start before optimization, rather than introducing a random functional change.

The final recipe used:

The serving evaluation uses 40 prompts held outside the complete training-source reservoir: 10 each from chat, code, math, and STEM. Both hardware matrices use concurrency 1, 4, and 8, greedy sampling, and up to 512 output tokens per request. The target is the official Ornith NVFP4 checkpoint, KV cache is FP8 E4M3, Marlin runs the MoE, and FlashInfer runs attention. The four matched configurations are autoregressive decoding, Ornith's built-in NEXTN head, the unadapted Qwen3.5 DFlash warm start, and this Ornith-adapted DFlash2 drafter. Both external drafts propose up to 10 tokens.

Ornith's NEXTN/MTP head is inherited from Qwen3.5 and was not trained during Ornith's continued pretraining, so it is not expected to be a strong speculative baseline. Its result is included as a reference for the head shipped with the target checkpoint.

The first matrix was measured on one NVIDIA DGX Spark.

The same matrix was measured on one NVIDIA H200 141GB HBM3e GPU.

On DGX Spark at concurrency 1, the draft configuration produced 114.2 output tokens/s, 44.0% above target-only serving, with mean accepted length 4.02 at concurrency 1. These measurements are specific to the recorded software, hardware, prompt mix, quantization, and serving settings; they are not general quality or speed claims. Machine-readable results are in spark-evaluation.json and h200-evaluation.json.

Use an SGLang build with DFlash2 support. The release evaluation used SGLang commit 710267dc4c817b38d9965346390cd56b59b54eda.

Memory limits, context length, and concurrency should be set for the deployment hardware rather than copied blindly from the evaluation setup.

The checkpoint is released under Apache-2.0. It was initialized from the Apache-2.0 Qwen3.5 DFlash checkpoint and trained against the MIT-licensed Ornith target. The training conversations derive from NVIDIA Nemotron Post-Training Dataset v2 under CC-BY-4.0. See the linked repositories for their notices and model cards.