# Ripple-65B Preview

> Source: https://aiwiki.ai/wiki/ripple_65b_preview
> Updated: 2026-11-06
> Categories: Large Language Models, Reasoning Models
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Ripple-65B Preview." aiwiki.ai, 6 Nov 2026. https://aiwiki.ai/wiki/ripple_65b_preview
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**Ripple-65B Preview** is an open-weight, text-based [large language model](https://aiwiki.ai/wiki/large_language_model) from [Ricursive Intelligence](https://aiwiki.ai/wiki/ricursive_intelligence). Released in November 2026, it is a general-purpose variant of a model the company uses in its chip-design platform. It uses a dense [Transformer](https://aiwiki.ai/wiki/transformers) architecture with about 65 billion parameters. The release includes instruction-tuned BF16 and experimental NVFP4 checkpoints, a base checkpoint, and a separate draft model for faster decoding.[1][2][3][4]

The weights are distributed under the [Apache License 2.0](https://aiwiki.ai/wiki/apache_license). Downloadable weights, a quantized checkpoint, and a hosted endpoint are different ways of using the model; their memory requirements and context limits depend on the serving configuration.[3][5][6]

## Release variants

The repositories belong to Ricursive's `ricursiveAI` organization on [Hugging Face](https://aiwiki.ai/wiki/hugging_face).[2][3][4]

| Repository | Contents |
| --- | --- |
| `ricursiveAI/Ripple-65B-Preview` | BF16 instruction-tuned model [2] |
| `ricursiveAI/Ripple-65B-Base-Preview` | BF16 pretraining checkpoint without midtraining or synthetic question-answer data [4] |
| `ricursiveAI/Ripple-65B-Preview-NVFP4` | Experimental instruction-tuned quantization: NVFP4 MLP layers and BF16 attention [3] |
| `ricursiveAI/Ripple-65B-Preview-DFlash2` | BF16 draft model that requires a Ripple target model [7] |

NVFP4 does not mean that every tensor uses four-bit storage. The released quantization configuration excludes attention modules, token embeddings and the output head from its four-bit linear-layer group.[8] In particular, the Hub's storage-derived parameter display for packed weights should not be substituted for the model's architectural parameter count.[3][8]

## Architecture

The released configuration names the model class `RicursiveForCausalLM`. It specifies 80 layers, a hidden size of 5,120, and 80 query heads sharing eight key/value heads. The model alternates four local-attention layers with one global-attention layer, giving 64 local and 16 global layers. The local sliding window is 1,024 tokens.[9]

| Configuration field | Released value |
| --- | --- |
| Layers | 80 |
| Hidden size | 5,120 |
| Feed-forward intermediate size | 40,960 |
| Query heads / key-value heads | 80 / 8 |
| Head dimension | 128 |
| Vocabulary size | 100,352 |
| Local attention window | 1,024 tokens |
| Configured maximum positions | 2,097,152 |

These values describe the configuration file, rather than a guarantee that a particular deployment can serve its maximum position count.[9]

The implementation applies rotary positional embeddings to [sliding-window attention](https://aiwiki.ai/wiki/sliding_window_attention) layers and omits them from the global layers. It also contains sigmoid-gated attention, per-head query/key normalization, value-residual connections, and normalization before and after attention and feed-forward operations.[10]

Ricursive cites prior research on gated attention, hybrid RoPE/NoPE attention and value residuals. Those papers study architectural components in their own experiments; their measured gains are not independent measurements of Ripple.[1][11][12][13]

## Training and development

Ricursive reports pretraining on 12 trillion tokens of predominantly English text and code. The instruction model then receives midtraining, long-context training, [supervised fine-tuning](https://aiwiki.ai/wiki/supervised_fine-tuning), long-context fine-tuning and mixed-domain [reinforcement learning](https://aiwiki.ai/wiki/reinforcement_learning).[2]

The company developed its recipe through smaller model-ladder experiments, starting from an [OLMo 3](https://aiwiki.ai/wiki/olmo_3) baseline. Its launch report describes changes to normalization, attention, data mixtures and learning-rate schedules. It reports up to a 16-fold improvement in code cross-entropy compute efficiency at the 4B scale. That measure concerns a particular training comparison, rather than a 16-fold improvement in application accuracy or inference speed.[1]

Ricursive puts the full development cost at $32 million, including experiments, synthetic data, evaluation, debugging and final training. Its comparison with other laboratories' costs is an estimate, not a disclosed accounting comparison on identical terms. It also reports reaching a Terminal-Bench 2.1 score of about 80.5 after 1.4 million reinforcement-learning rollouts.[1]

NVIDIA states that Ripple was trained on B300 GPUs and used open Nemotron post-training datasets.[14] The model-ladder approach has a separate research lineage: Bhagia and colleagues studied predicting task performance using smaller training runs. That paper predates Ripple and does not validate Ricursive's reported cost or efficiency figures.[15]

## Benchmark results

The following are **Ricursive-reported BF16 instruction-model results**, not an independent overall intelligence ranking.[2]

| Benchmark | Reported result |
| --- | --- |
| Terminal-Bench 2.1, Terminus 2 | 80.5 |
| SWE-bench Verified | 76.5 |
| SWE-bench Pro v1.0 | 61.42 |
| SWE-bench Pro v2.0 Hard | 79.1 |
| DeepSWE 1.1 | 44.3 |
| GPQA Diamond | 85.2 |
| AIME 2026 | 95.0 |
| AA-LCR | 75.0 |

Ricursive's comparison combines official competitor scores with in-house measurements. It identifies Terminus 2 for its SWE-bench Pro v2.0 Hard and DeepSWE runs, and otherwise OpenCode where no harness is specified. The card acknowledges substantial harness and context sensitivity on [Terminal-Bench](https://aiwiki.ai/wiki/terminal_bench).[2]

### Independent data-analysis evaluation

Plotly published its own evaluation on November 5, 2026, using a private synthetic wind-energy dataset and its Plotly Studio Embedded analytics harness. Ripple passed 35 of 43 questions. It answered all five statistics-trap questions and rejected all five requests unsupported by the data, but found none of five deliberately hidden patterns. Plotly also documented a fabricated numerical table following a failed query.[16]

Plotly reported $2.21 for the completed exam and a median 128 seconds per question under the provider prices and caching behavior used for that run. It restarted part of the exam after an unfinished query; the abandoned attempt was excluded from the reported cost and timing. The article describes one sample per model and notes that repeat runs of other models varied by up to three questions. These results describe that analytics task and harness, rather than all coding or reasoning workloads.[16]

## Context length and deployment

The BF16 card advertises a 1Mi-token [context window](https://aiwiki.ai/wiki/context_window), or 1,048,576 tokens. This should be distinguished from the configuration's 2,097,152 maximum positions and from a runtime's selected limit.[2][9]

| Deployment material | Documented configuration |
| --- | --- |
| BF16 B200 server | Two B200 GPUs, SGLang 0.5.19, BF16 target and KV cache, 1,048,576-token target context including output [6] |
| NVFP4 local launcher | DGX Spark/GB10, one running request, BF16 KV cache, 131,072-token context [17] |
| DFlash2 checkpoint | Separate drafter; configured maximum positions of 131,072 [7] |

The B200 launcher keeps a 16,384-token draft window even when the target accepts a longer context. Extending the draft's positional cache does not enlarge that attention window.[6]

Ricursive's server exposes separate reasoning text, final-answer text and structured tool calls. Applications execute requested tools and return their results to the model. The supplied template has no reasoning-effort levels or thinking on/off switch. The server documentation also identifies stricter function-schema requirements for required tool calls in its tested SGLang version.[6]

### Local quantization and storage

The NVFP4 repository contains six weight shards totaling 57,213,852,024 bytes, approximately 57.2 GB in decimal units. The BF16 repository's 16 shards total 129,565,519,472 bytes, approximately 129.6 GB. These are file-size totals from the repository inventories, not measured peak RAM or GPU-memory use; inference also needs cache, activations and runtime buffers.[18][19]

NVIDIA promotes the NVFP4 checkpoint for [DGX Spark](https://aiwiki.ai/wiki/nvidia_dgx_spark), and the official local launcher targets that machine. Its 131,072-token setting does not establish that a single Spark can run the advertised 1Mi context with the same configuration.[14][17]

### DFlash2 draft model

DFlash2 is an auxiliary model for [speculative decoding](https://aiwiki.ai/wiki/speculative_decoding), rather than a second standalone assistant. The target checks its proposed tokens. Correct rejection sampling is required to preserve the target's sampling distribution.[7]

The draft card reports 4,037,752,064 stored BF16 parameters, excluding shared embeddings and the output head supplied by the target. It has five layers and drafts 16-token blocks. Its recorded export validation used a particular `ripple-sglang` runtime, eight H100 GPUs and an RL-step-500 target checkpoint. The card explicitly limits what that validation establishes for other runtimes and target revisions.[7]

## License and limitations

Apache 2.0 permits use, modification and redistribution subject to its conditions, including retaining required notices. It does not grant trademark rights, and the released materials carry the license's warranty disclaimer.[5]

Ripple is a preview release. Its published coding results, Plotly's analytics results and a deployment's measured behavior answer different questions. BF16 benchmark scores should not be assumed to describe the experimental NVFP4 checkpoint, and a configured long context should not be read as evidence of uniform retrieval accuracy across every input length.[2][3][16]

## References

1. Ricursive Intelligence. [Introducing Ripple-65B Preview](https://www.ricursive.com/ripple). Launch report, November 2026.
2. Ricursive Intelligence. [Ripple-65B Preview model card](https://huggingface.co/ricursiveAI/Ripple-65B-Preview). Accessed November 7, 2026.
3. Ricursive Intelligence. [Ripple-65B Preview NVFP4, experimental model card](https://huggingface.co/ricursiveAI/Ripple-65B-Preview-NVFP4). Accessed November 7, 2026.
4. Ricursive Intelligence. [Ripple-65B Base Preview model card](https://huggingface.co/ricursiveAI/Ripple-65B-Base-Preview). Accessed November 7, 2026.
5. Ricursive Intelligence. [Ripple checkpoint Apache License 2.0](https://huggingface.co/ricursiveAI/Ripple-65B-Preview/blob/da94cc013a8fda04a502f0a185083c7099ea06c1/LICENSE).
6. Ricursive Intelligence. [Ripple on B200: server documentation](https://huggingface.co/ricursiveAI/Ripple-65B-Preview/blob/da94cc013a8fda04a502f0a185083c7099ea06c1/inference/server/README.md).
7. Ricursive Intelligence. [Ripple-65B Preview DFlash2 model card](https://huggingface.co/ricursiveAI/Ripple-65B-Preview-DFlash2). Accessed November 7, 2026.
8. Ricursive Intelligence. [NVFP4 checkpoint configuration](https://huggingface.co/ricursiveAI/Ripple-65B-Preview-NVFP4/blob/99eea4477a001a41b86914e41feb4b2ed86105e5/config.json).
9. Ricursive Intelligence. [BF16 checkpoint configuration](https://huggingface.co/ricursiveAI/Ripple-65B-Preview/blob/da94cc013a8fda04a502f0a185083c7099ea06c1/config.json).
10. Ricursive Intelligence. [Ricursive Transformer implementation](https://huggingface.co/ricursiveAI/Ripple-65B-Preview/blob/da94cc013a8fda04a502f0a185083c7099ea06c1/modeling_ricursive.py).
11. Zihan Qiu and colleagues. [Gated Attention for Large Language Models: Non-linearity, Sparsity, and Attention-Sink-Free](https://arxiv.org/abs/2505.06708). May 10, 2025.
12. Bowen Yang and colleagues. [Rope to Nope and Back Again: A New Hybrid Attention Strategy](https://arxiv.org/abs/2501.18795). Revised October 22, 2025.
13. Zhanchao Zhou and colleagues. [Value Residual Learning](https://arxiv.org/abs/2410.17897). Revised June 8, 2025.
14. NVIDIA RTX Spark. [Announcement of Ripple training and local NVFP4 deployment](https://x.com/NVIDIARTXSpark/status/2118071270965407879). November 2026.
15. Akshita Bhagia and colleagues. [Establishing Task Scaling Laws via Compute-Efficient Model Ladders](https://arxiv.org/abs/2412.04403). COLM 2025, revised August 22, 2025.
16. Plotly. [Ricursive Ripple-65B Preview Benchmark Results for Data Analysis](https://plotly.com/blog/ricursive-ripple-65b-preview-plotly-data-analytics-bench/). November 5, 2026.
17. Ricursive Intelligence. [DGX Spark local launcher](https://huggingface.co/ricursiveAI/Ripple-65B-Preview-NVFP4/blob/99eea4477a001a41b86914e41feb4b2ed86105e5/inference/local/spark.py).
18. Hugging Face. [BF16 repository file metadata](https://huggingface.co/api/models/ricursiveAI/Ripple-65B-Preview?blobs=true). Snapshot inspected November 7, 2026.
19. Hugging Face. [NVFP4 repository file metadata](https://huggingface.co/api/models/ricursiveAI/Ripple-65B-Preview-NVFP4?blobs=true). Snapshot inspected November 7, 2026.