Ripple 65B Preview Release - Base - GSM8K (CoT EM) - 8-shot: leaderboard
Ricursive Ripple-65B Preview native base comparison, immutable card revision e27e171792f303d74c048184e8259ba663cfa0d5. Pretrain-only base checkpoints; no instruct or midtraining merge. Exact native metric and shot count are in the board name; pass@1 sample counts and temperatures remain as printed. Official-card and internal Olmes-fork results retain distinct provenance. Native percentages, higher is better. Ripple-65B Base Preview:in-house/Ricursive Olmes fork; Laguna M.1 225B:in-house/Ricursive Olmes fork; DeepSeek V3.1 671B:in-house/Ricursive Olmes fork; DeepSeek V4 Flash 284B:official-card; Nemotron 3 Ultra 550B:official-card. Terminal-Bench self-reported and Terminus2 comparisons are separate; native table80.5 is retained, not replaced by prose80.52. No throughput/cost/model-size row is a benchmark. This source family alone does not establish independent main-rank eligibility.
Metric: Native GSM8K (CoT EM) (%). Source: huggingface.co. Saturation forecast: Rough model projection: around 2026. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek V4 Flash 284B (Ripple Release Run: phase=Base;origin=official-card;precision=unreported;shots=8-shot) | 90.8 | |
| 2 | Ripple-65B Base Preview (Ripple Release Run: phase=Base;origin=in-house;precision=BF16;shots=8-shot;harness=Ricursive Olmes fork) | 88.5 | |
| 3 | Nemotron 3 Ultra 550B (Ripple Release Run: phase=Base;origin=official-card;precision=unreported;shots=8-shot) | 88.1 | |
| 4 | Laguna M.1 225B (Ripple Release Run: phase=Base;origin=in-house;precision=unreported;shots=8-shot;harness=Ricursive Olmes fork) | 80.4 | |
| 5 | DeepSeek V3.1 671B (Ripple Release Run: phase=Base;origin=in-house;precision=unreported;shots=8-shot;harness=Ricursive Olmes fork) | 79.6 |
Interactive version: theaggregate.ai/benchmark?slug=ripple-65b-preview-release-base-gsm8k-cot-em-8-shot · How It Works · Data refreshed daily, snapshot 2026-11-06.