ricursiveAI/Ripple-65B-Preview-DFlash2

Ripple-65B Preview DFlash2

BF16 | NVFP4 | Base Model | DFlash2 | Launch Blog | License: Apache 2.0

Ripple-65B Preview DFlash2 is a draft model for speculative decoding with Ripple-65B Preview or Ripple-65B Preview NVFP4. This repo contains the BF16 drafter. It requires the target model to generate text.

The drafter proposes multiple tokens at once, conditioned on intermediate hidden states from Ripple. The target then verifies the proposals. A serving implementation with correct speculative rejection sampling preserves the target model's sampling distribution.

Model

DFlash2 combines parallel block drafting, dynamic convolutions, and a candidate selector. This checkpoint drafts blocks of 16 tokens, retains 32 candidates per position, and uses a rank-256 selector. Its dynamic convolutions have kernel size 4 and group size 16.

The draft has four sliding-attention layers with a 4,096-token window, followed by a full-attention layer without rotary position embeddings (NoPE). The full-attention layer can attend beyond the sliding window.

The checkpoint contains 4.04B parameters (4,037,752,064), all stored in BF16. This count is derived from the tensors in model.safetensors; it excludes the target model's shared token embedding and output head, which are not stored in this repository.

Property	Value
Architecture	DFlash2DraftModel
Draft layers	5
Hidden size	5,120
Query / KV heads	80 / 8
Head dimension	128
MLP	SwiGLU, intermediate size 40,960
Attention pattern	4 sliding-attention layers + 1 full-attention layer
Sliding window	4,096 tokens
NoPE layer	Draft layer 4 (zero-based)
RoPE theta	10,000 on rotary-enabled layers
Target hidden-state layers	1, 20, 39, 58, 77 (zero-based)
Target depth	80 layers
Vocabulary size	100,352
Draft block size	16
Dynamic convolution	Kernel size 4, group size 16
Candidate selector	Top-k 32, rank 256
Configured maximum positions	131,072
Weight precision	BF16

Serve

Load this repository as the speculative draft model alongside Ripple-65B Preview. Use the target model's tokenizer, chat template, token embeddings, and output head.

The serving runtime must support Ripple and DFlash2DraftModel, including this checkpoint's per-layer attention pattern, NoPE layer, four-tap convolutions, and target hidden-state taps. Preserve these settings from config.json when loading the draft.

For SGLang-based serving, the draft model settings are:

--speculative-algorithm DFLASH \
--speculative-draft-model-path ricursiveAI/Ripple-65B-Preview-DFlash2 \
--speculative-draft-model-quantization unquant \
--speculative-num-draft-tokens 16

These flags supplement a Ripple-compatible server launch; they are not a standalone launch command. The export's recorded validation used the ripple-sglang runtime on eight H100 GPUs with the RL-step-500 target checkpoint. That validation does not establish compatibility or performance for other target revisions or runtime versions.

Citation

@misc{ricursive2026ripple65b,
      title={Ripple-65B: Pushing the Frontier of Local Intelligence},
      author={Ricursive Intelligence},
      year={2026},
}

Contact

For questions and comments, please contact us at info@ricursive.com. To learn more, please visit our website.