Overview
Ripple-65B Preview is the first open-weight model from Ricursive Intelligence, published on Hugging Face on November 4, 2026 under the Apache 2.0 license. It is a 65-billion-parameter dense Transformer pretrained from scratch on 12 trillion tokens of text-only, predominantly English and code data, with support for up to 1M tokens of context. The Preview release includes the instruction-tuned model in BF16 and NVFP4 numerics and a BF16 base model.
Ricursive chose a dense model rather than a Mixture-of-Experts to keep pretraining and post-training infrastructure simple. The model uses a thin-and-deep shape — a narrow d_model of 5120 and 80 layers — and a Gemma 3N-like architecture with overprovisioned attention heads, a higher sliding-window-to-global layer ratio with a smaller window, FFN activation clipping, gated attention, global NoPE and value residuals. Ricursive says training started from an Olmo 3 baseline and ran at about 40% MFU in BF16 with no custom kernels or quantization.
Post-training followed a standard recipe of midtraining (200B tokens at 64K), long-context training at 2M sequence length, SFT (183B tokens), long-context SFT and mixed-domain RL, with a focus on agentic coding; Ricursive describes the Preview's post-training as relatively light.
In Ricursive's model card the instruct model scores 80.5 on Terminal-Bench 2.1, 76.5 on SWE-bench Verified, 79.1 on SWE-bench Pro v2.0 Hard, 85.2 on GPQA Diamond and 98.3 on AIME 2025, set against Nemotron 3 Ultra 550B, Inkling 975B, Muse Spark Open 816B and Reflection Beam 501B. Ricursive notes that some competitor cells come from its own in-house eval runs where official scores were missing, and that the model does not lead on every benchmark.
Ricursive says the model fits on consumer hardware such as the NVIDIA DGX Spark, provides an NVFP4 checkpoint for local use, and reaches 250 tokens per second on two B200 GPUs at batch size 8 before quantization or custom kernels.
| Released | 2026-11-04 |
|---|---|
| License | Apache-2.0 |
| Weights | Open weights |
| Parameters | 65B (dense) |
| Context | Up to 1M tokens |
| Architecture | Dense Transformer, 80 layers, d_model 5120 (Gemma 3N-like) |
| Modalities | Text |
| Status | Preview. Open weights on Hugging Face (Apache 2.0): the instruction-tuned model in BF16 and NVFP4, a BF16 base model and a DFlash2 variant. |
Benchmarks
Ripple-65B Preview against US open-weight models, transcribed from Ricursive's model card. Ricursive shows official scores where available and fills missing cells with its own in-house evals; all models tested at BF16 with maximum effort.
| Benchmark | Ripple-65B Preview | Nemotron 3 Ultra 550B | Inkling 975B | Muse Spark Open 816B | Reflection Beam 501B |
|---|---|---|---|---|---|
| MMLU | 89.6 | 92 | 91.3 | 92.6 | 92.4 |
| MMLU-Pro | 84.2 | 86.8 | 87.3 | 87.9 | 76.5 |
| GPQA Diamond | 85.2 | 87 | 87.2 | 89 | 90.6 |
| AIME 2024 | 97.1 | 94.2 | 97.5 | 97.9 | 96.7 |
| AIME 2025 | 98.3 | 94.2 | 91.7 | 100 | 99.6 |
| AIME 2026 | 95 | 95.4 | 97.1 | 98.3 | 97.7 |
| HMMT Feb 2026 | 90.5 | 80.3 | 86.3 | 93.9 | 93.9 |
| Terminal-Bench 2.1 (self-reported) | 80.5 | 56.4 | 63.8 | 76.4 | 80.1 |
| Terminal-Bench 2.1 (Terminus 2) | 80.5 | 55.1 | 56.2 | 78.7 | 70.8 |
| SWE-bench Verified | 76.5 | 71.9 | 77.6 | 78.7 | 81.6 |
| SWE-bench Pro v1.0 | 61.42 | 46.9 | 54.3 | 58.2 | 66.1 |
| SWE-bench Pro v2.0 Hard | 79.1 | 30.7 | 2 | 52.3 | 76.5 |
| DeepSWE 1.1 | 44.3 | 0.2 | 1.8 | 60 | 43.3 |
| SciCode (Inspect AI) | 39.4 | 44.6 | 46.1 | 58.5 | 49.2 |
| τ³-bench banking | 31.27 | 22.6 | 23.7 | 14.4 | 36.1 |
| AutomationBench (Public) | 36.4 | 8.4 | 14.6 | 42.4 | 39.3 |
| MCP Atlas | 71.67 | 62.2 | 76 | 89.2 | 80.8 |
| FORTRESS Adversarial | 83.8 | 79.3 | 78 | 94.2 | 92.1 |
| FORTRESS Benign | 95.5 | 89.1 | 95.9 | 82.6 | 86.4 |
| StrongREJECT | 97.5 | 98.2 | 98.6 | 94.7 | 95.8 |
| IFBench | 79.2 | 81.7 | 79.8 | 78.6 | 79.6 |
| AA-LCR | 75 | 65.4 | 78.3 | 80.7 | 78.2 |
This model's scores
- AIME 202598.3%
- MMLU89.6%
- GPQA Diamond85.2%
- Terminal-Bench 2.180.5%
- SWE-bench Pro v2.0 Hard79.1%
- SWE-bench Verified76.5%
- AA-LCR75%
- MCP Atlas71.67%
- SWE-bench Pro v1.061.42%
- DeepSWE 1.144.3%
Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.
Strengths
- 65B dense model — no Mixture-of-Experts routing — with up to 1M tokens of context
- 80.5 on Terminal-Bench 2.1 and 79.1 on SWE-bench Pro v2.0 Hard in Ricursive's model card
- Apache 2.0 weights in BF16 and NVFP4, plus a BF16 base model
- Small enough for an NVIDIA DGX Spark, per Ricursive
Best for
- Reach for it for self-hosted agentic coding and terminal work where a 65B model fits your hardware.
- Reach for it as a permissively licensed base model for further training, using the Ripple-65B Base Preview checkpoint.
- Look elsewhere for image, audio or video input: Ripple-65B Preview is a text-only model.
How to access
| Provider | Model ID |
|---|---|
| OpenRouter ↗ | — |
FAQ
What is Ripple-65B Preview?
Ripple-65B Preview is Ricursive Intelligence's first open-weight language model, published on Hugging Face on November 4, 2026. It is a 65B-parameter dense Transformer pretrained on 12 trillion tokens, tuned for agentic coding, with support for up to 1M tokens of context.
What license is Ripple-65B Preview released under?
Apache 2.0. The release includes the instruction-tuned model in BF16 and NVFP4, a BF16 base model (a pretrain-only checkpoint) and a DFlash2 variant on Hugging Face.
How does Ripple-65B Preview score on coding benchmarks?
In Ricursive's model card it scores 80.5 on Terminal-Bench 2.1, 76.5 on SWE-bench Verified, 61.42 on SWE-bench Pro v1.0, 79.1 on SWE-bench Pro v2.0 Hard and 44.3 on DeepSWE 1.1. On SWE-bench Verified the same table lists 81.6 for Reflection Beam 501B. Some competitor scores are Ricursive's own in-house runs.
Can Ripple-65B Preview run locally?
Ricursive says the 65B model fits on consumer hardware such as the NVIDIA DGX Spark and publishes an NVFP4 checkpoint and a Colab notebook to help with local use.