Ripple-65B-Preview — SAGE REXL3 2.57 bpw
Loadable ExLlamaV3 / REXL3 pack of ricursiveAI/Ripple-65B-Preview.
Quantized by vcruz305 with SAGE, our own
mixed-precision quantization method for EXL3. Tensors use different EXL3 K
values; the kernels read K from each trellis tensor at load.
~20.8 GiB GPU-resident weights (22.3 GB pack, text-only). This 2.57 bpw REXL3 quantization preserves high token-level agreement (96.49% Top-1, 0.0155 mean KLD) with Ricursive's official BF16 release (ricursiveAI/Ripple-65B-Preview).
Measured quality
Teacher-forced token-level agreement with Ricursive's official BF16 release (ricursiveAI/Ripple-65B-Preview), full-vocabulary KL divergence per token, on held-out sequences (assistant-turn tokens, 55,535 scored):
| Pack | Resident footprint | KL vs official BF16 release (mean) | KL vs official BF16 release (p50) |
Top-1 Agreement | Top-5 Agreement | Top-5 Overlap |
|---|---|---|---|---|---|---|
| this pack, 2.57 bpw | ~20.8 GiB | 0.0155 | 0.000068 | 96.49% | 99.88% | 85.32% |
| Official BF16 release | native | 0 | 0 | 100% | 100% | 100% |
Per-category breakdown
| Category | KL vs official BF16 release (mean) | KL vs official BF16 release (p50) |
Top-1 Agreement | Answer-Only Top-1 | Top-5 Overlap |
|---|---|---|---|---|---|
| general | 0.0078 | 0.000014 |
97.62% | – | 85.25% |
| prose_narrative | 0.0182 | 0.000441 |
95.64% | – | 86.72% |
| multilingual | 0.0436 | 0.001424 |
94.03% | – | 83.26% |
| reasoning | 0.0072 | 0.000015 |
98.15% | – | 86.50% |
| code | 0.0068 | 0.000048 |
98.64% | – | 87.26% |
| ifeval | 0.0025 | 0.000014 |
98.46% | – | 82.17% |
| gpqa_stem | 0.0394 | 0.009184 |
89.84% | – | 88.39% |
| swe_patch | 0.0059 | 0.000030 |
98.44% | – | 86.82% |
| math | 0.0065 | 0.000053 |
96.83% | – | 83.26% |
| cyber | 0.0089 | 0.000017 |
98.22% | – | 84.43% |
| mmlu_pro | 0.0053 | 0.000017 |
97.15% | – | 82.69% |
Held-out & long-context suites (assistant-turn tokens)
| Suite | Context | Assistant-turn tokens | KL vs official BF16 release (mean) | KL vs official BF16 release (p50) |
Top-1 Agreement | Top-5 Agreement | Answer-Only Top-1 | Top-5 Overlap |
|---|---|---|---|---|---|---|---|---|
| wiki1100 | 1,100-token prefix | 890 | 0.0674 | 0.020488 |
89.66% | 99.55% | – | 86.72% |
| long17k | 17,408-token long context | 10,654 | 0.0412 | 0.014641 |
91.45% | 99.83% | – | 87.86% |
Pack
| Item | Value |
|---|---|
| Format | EXL3 / REXL3 (trellis) |
| Profile | 2.57 bpw SAGE mixed-K |
| Resident footprint | ~20.8 GiB (text-only) |
| Shards | 10 (model-00001-of-00010.safetensors .. model-00010-of-00010.safetensors) |
| Index | model.safetensors.index.json |
quant_method |
exl3 |
model_type |
ricursive |
Serve
Clone vcruz305/Ripple-65B-Preview-EXL3-SGLang-recipe and serve with SGLang 0.5.19 + vcruz305/sglang-exl3 and the quantized DFlash2 draft (vcruz305/Ripple-65B-Preview-DFlash2-SAGE-REXL3):
bash scripts/setup.sh
MODEL_DIR=/path/to/Ripple-65B-Preview-SAGE-REXL3-2.57bpw bash scripts/serve.sh
Load
ExLlamaV3, point at this repo (or a local snapshot):
from exllamav3 import Config, Model, Tokenizer
cfg = Config.from_directory("vcruz305/Ripple-65B-Preview-SAGE-REXL3-2.57bpw")
model = Model.from_config(cfg)
model.load()
tokenizer = Tokenizer.from_config(cfg)
About SAGE
SAGE is our own mixed-precision quantization method for EXL3. The method is not published. If a tensor is EXL3, its K is self-describing on the trellis.
Source model (Ricursive, not this pack)
Ripple-65B-Preview is Ricursive Intelligence's dense 65B autoregressive model: 80 layers, hidden size 5120, GQA with 80 query / 8 KV heads, gated attention, value-residual connections, and a hybrid sliding-window / full-attention pattern (sliding window 1024; full attention every 5th layer) with 2,097,152 max positions. Architecture, sampling defaults, and any benchmark rows on the upstream card are Ricursive's.
License
Apache 2.0, same as Ripple-65B-Preview. Cite Ricursive for the base model. This repository is the EXL3 pack only.