Ripple-65B PreviewA 65B dense Transformer pretrained from scratch on 12 trillion tokens, with 1M context and an Apache 2.0 license. Released November 4, 2026. What it is, which checkpoint to download, how to run it, and what the benchmarks actually say — including the weak spots.
Ripple-65B Preview is the first open-weight model from Ricursive Intelligence (repository ricursiveAI/Ripple-65B-Preview), a 65-billion-parameter dense model — not a mixture-of-experts — trained from scratch on a 12T-token text and code corpus. It uses a thin-and-deep architecture (d_model 5120, 80 layers), starting from an Olmo 3 baseline and switching to a Gemma 3N-like architecture for training stability.
The headline claim is agentic coding: 80.52 on Terminal-Bench 2.1, ahead of much larger US open-weight models (Reflection Beam 501B at 80.1, Meta Muse Spark Open 816B at 76.4). Ricursive says the full R&D cost was $32M. The "Preview" label matters: post-training was relatively light, with a focus on agentic coding — this is not a finished flagship.
| Repository | ricursiveAI/Ripple-65B-Preview (BF16 instruct) |
|---|---|
| Size | 129.6 GB (129,565,519,472 bytes), 16 safetensors shards |
| Parameters | 64,782,689,360 (BF16) |
| Architecture | Dense Transformer, RicursiveForCausalLM (custom code) |
| Context | 1,048,576 tokens (1M) |
| Training | 12T tokens pretrain; midtraining, SFT, long-context SFT, RL |
| License | Apache 2.0 — no gated access, no approval needed |
| Other versions | NVFP4 instruct (ricursiveAI/Ripple-65B-Preview-NVFP4), BF16 Base (pretrain-only), DFlash2 variant — all linked from the model card |
Three official checkpoints. Pick by your hardware:
| Checkpoint | Download size | VRAM need | Best for |
|---|---|---|---|
Ripple-65B-Preview (BF16 instruct) | 129.6 GB, 16 shards | Highest (~130 GB weights alone) | Max quality, datacenter GPUs |
Ripple-65B-Preview-NVFP4 (NVFP4 instruct, experimental) | 101.0 GB, 6 shards | Lower — vendor says it fits NVIDIA DGX Spark | Best balance — local / single-machine runs |
Ripple-65B-Base-Preview (BF16 base) | ~129.6 GB | Highest | Fine-tuning and research (pretrain-only, no instruction tuning) |
Why is the NVFP4 download still 101 GB? It is a conservative quantization: only the MLP layers are NVFP4, attention stays in BF16. It is not a full 4-bit model — expect a real but modest size saving, not a 4x shrink.
No approval or gated access — the repositories are public. Install huggingface_hub first (pip install -U huggingface_hub), then:
hf download ricursiveAI/Ripple-65B-Preview \
--local-dir ./Ripple-65B-PreviewOr the NVFP4 version for local hardware:
hf download ricursiveAI/Ripple-65B-Preview-NVFP4 \
--local-dir ./Ripple-65B-Preview-NVFP4Make sure you have the disk space first: ~130 GB free for BF16, ~101 GB for NVFP4. No account, token, or waitlist is needed. If you only want to try it first, the model page lists a free ZeroGPU demo — no download required.
This model uses a custom architecture (RicursiveForCausalLM). The repository ships configuration_ricursive.py and modeling_ricursive.py, and Hugging Face tags it custom_code. Standard loading without the trust flag will fail:
from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("ricursiveAI/Ripple-65B-Preview")
model = AutoModelForCausalLM.from_pretrained(
"ricursiveAI/Ripple-65B-Preview",
trust_remote_code=True,
torch_dtype="bfloat16",
device_map="auto")The repository also ships its own inference code — inference/server/serve.py for serving and inference/local/spark.py plus a notebook for DGX Spark. vLLM support for this custom architecture is not documented in the model card, so check compatibility before assuming your usual vLLM flow works.
trust_remote_code=True. The second is disk space: the BF16 download is 129.6 GB, and you need headroom on top of that.| Checkpoint | Weights on disk | What runs it |
|---|---|---|
| BF16 instruct | 129.6 GB | Datacenter GPUs. Vendor benchmarked 250 tok/s on 2× B200 (batch 8, before quantization). |
| NVFP4 instruct | 101.0 GB | Vendor says it fits on NVIDIA DGX Spark (128 GB unified memory). Exact runtime VRAM is not published. |
| BF16 base | ~129.6 GB | Same as BF16 instruct — datacenter GPUs. |
VRAM figures above are weights on disk (verified) and vendor claims, not measured runs. Budget extra memory for KV cache at 1M context — long-context inference is where memory actually blows up.
Every number below is from the model's own card, reported by Ricursive or its in-house eval flow — none of it is independently verified, and the card itself warns that "replicating benchmark scores has been challenging." Rival scores in the table are Ricursive's own runs, not the other labs' numbers.
| Benchmark | Ripple-65B Preview | Best rival in the table |
|---|---|---|
| Terminal-Bench 2.1 (self-reported) | 80.52 | Reflection Beam 501B — 80.1 |
| Terminal-Bench 2.1 (Terminus 2 harness) | 80.5 | Muse Spark Open 816B — 78.7 |
| SWE-bench Verified | 76.5 | Reflection Beam 501B — 81.6 |
| SWE-bench Pro v2.0 Hard | 79.1 | Reflection Beam 501B — 76.5 |
| DeepSWE 1.1 | 44.3 | Muse Spark Open 816B — 60.0 |
| MMLU / MMLU-Pro | 89.6 / 84.2 | Muse Spark Open 816B — 92.6 / 87.9 |
| GPQA Diamond | 85.2 | Reflection Beam 501B — 90.6 |
| τ³-bench banking | 31.27 | Reflection Beam 501B — 36.1 |
The pattern is clear: Ripple-65B leads on agentic coding (Terminal-Bench, SWE-bench Pro Hard) and trails on knowledge, math, and tool-use breadth. It is a coding specialist, not an all-rounder — and with light post-training, the "Preview" label is doing real work.
Yes. Apache 2.0 allows download, fine-tuning, deployment, and selling products built on it, with no copyleft. Standard Apache 2.0 attribution and license-notice conditions apply.
No. Ricursive Intelligence is a separate company; "Ripple" here is the model name. Ripple Labs, Inc. owns the RIPPLE trademark for its payments business. Do not confuse the two.
The model card only compares against US open-weight models, so Kimi and DeepSeek are not in its table. Kimi K3.1 Pro's weights were released the same day (November 4, 2026) — a much larger MoE model; Ripple-65B's pitch is doing more with 65B dense parameters, not beating everything.
No date has been announced. The Preview is fully usable and Apache-licensed; if a final version ships, it will be a new repository, not an in-place update.
Only the MLP layers are quantized to NVFP4 — attention, embeddings, and the language-model head stay in BF16. That is why the download is 101 GB instead of ~35 GB. It is labeled experimental for the same reason.
Yes — the model page lists a free ZeroGPU demo. It is the fastest way to check whether the model's coding style fits your use case before committing disk space.