Ripple-65B Preview

A 65B dense, text-only model from a chip-design lab. Apache 2.0, strong on agentic coding, with an NVFP4 build for a DGX Spark.

In short. Ripple is a 65-billion-parameter text model from Ricursive Intelligence, a US startup that uses AI to design chips. It is aimed at coding agents, and the Apache 2.0 license allows commercial use. At full precision it needs about 130 GB of memory, and a 57 GB NVFP4 build is meant for desktop machines such as the NVIDIA DGX Spark.

Summary

Ricursive Intelligence, a US lab that builds AI for chip design, released Ripple-65B Preview on 4 November 2026 under the Apache 2.0 license. Ricursive calls it a general-purpose variant of the proprietary model behind its chip-design platform. It is a text-only model with 64.8 billion parameters, pretrained from scratch on 12 trillion tokens of mostly English text and code.

Ripple is dense, not a mixture of experts, so every parameter is used for every token. Ricursive says that choice kept pretraining and post-training simple enough for a team of one to five people, at about 40% hardware utilization and with no custom kernels. The network is thin and deep: a hidden size of 5,120, 80 layers, 80 attention heads and 8 key-value heads. Sixty-four layers attend over a sliding window of 1,024 tokens and 16 use global attention without positional encoding, which keeps long-context memory in check. The recipe started from the Olmo 3 baseline from AI2 and added FFN activation clipping, gated attention and value residuals. Pretraining used 8K-token sequences, followed by 200 billion midtraining tokens, a long-context stage with sequences of up to 2 million tokens, supervised fine-tuning and mixed-domain reinforcement learning. Ricursive describes the post-training as relatively light and focused on agentic coding. The card promises a context of 1,048,576 tokens, while the configuration allows 2,097,152 positions. Ricursive also puts its full research and development cost at US$32 million, which is its own estimate.

Ricursive reports 80.5 on Terminal-Bench 2.1, ahead of the 80.1 it lists for Reflection Beam, 76.4 for Muse Spark Open, 63.8 for Inkling and 56.4 for Nemotron 3 Ultra. Its own table also shows where others lead: Beam scores higher on SWE-bench Verified (81.6 against 76.5), and Muse Spark Open scores higher on MCP Atlas (89.2 against 71.7). Ricursive notes that agentic scores shift with the test harness, and several rival numbers in the table are its own in-house runs. Four repositories are published: the BF16 instruction-tuned model (129.6 GB), a pretrain-only BF16 base model, an experimental NVFP4 build of 57.2 GB that quantizes only the MLP layers and keeps attention in BF16, and a speculative-decoding draft called DFlash2. The card says the NVFP4 build fits on an NVIDIA DGX Spark, and a Colab notebook is included. For servers, the repository ships an SGLang recipe for two B200 GPUs, where Ricursive reports 250 tokens per second at batch size 8 before quantization or custom kernels.

Good for

  • Agentic coding and terminal work on one multi-GPU node or a high-memory workstation, instead of a rack of eight GPUs
  • Teams that want a US-built model under Apache 2.0, with no revenue conditions
  • Local experiments on a 128 GB desktop machine using the 57 GB NVFP4 build
  • Research on dense versus mixture-of-experts designs, since a base model and a speculative-decoding draft are published too

Watch out for

  • It is a preview, and Ricursive describes the post-training as relatively light and focused on agentic coding
  • It handles text only, mostly English and code, with no image input
  • The benchmark numbers are self-reported, several rival scores are Ricursive's own runs, and Terminal-Bench scores moved with the harness
  • All 64.8 billion parameters run for every token, so compute per token is higher than for a mixture-of-experts model with about 23 billion active
  • The NVFP4 build is marked experimental, and the 1M-token context is the developer's claim that we have not checked

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Estimates for the Ripple-65B Preview variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.

Variants Parameters BF16 FP8 / INT8 4-bit (Q4) Fits on
Ripple-65B Preview BF16 64.8B 149 GB 75 GB 45 GB Workstation
Ripple-65B Preview NVFP4 (experimental) NVFP4 64.8B 149 GB 75 GB 63 GB Workstation
Ripple-65B Base Preview BF16 64.8B 149 GB 75 GB 45 GB Workstation

Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.

VRAM calculator

The license in plain words

One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.

More about Apache 2.0

See also

Frequently asked questions

What is Ripple-65B Preview?
Ripple is a 65-billion-parameter text model from Ricursive Intelligence, a US startup that uses AI to design chips. It is aimed at coding agents, and the Apache 2.0 license allows commercial use. At full precision it needs about 130 GB of memory, and a 57 GB NVFP4 build is meant for desktop machines such as the NVIDIA DGX Spark.
Can I use Ripple-65B Preview commercially?
Yes. Ripple-65B Preview is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
Is Ripple-65B Preview open source?
Ripple-65B Preview is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
Can I run Ripple-65B Preview on my own computer?
Only on a workstation. The smallest variant, Ripple-65B Preview, needs about 45 GB of memory at 4-bit precision, so plan for a 48 to 96 GB GPU, several graphics cards, or a Mac with plenty of memory.
When was Ripple-65B Preview released?
Ripple-65B Preview was released by Ricursive Intelligence on November 4, 2026.
How long a context window does Ripple-65B Preview have?
Ripple-65B Preview supports a context window of up to 1M tokens.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
Reflection Beam Reflection AI 501B 23.4B active 1M Apache 2.0 (Permissive) 323 GB Big-memory desktop
Nanbeige4.3-33B Nanbeige LLM Lab 33.1B 512K Apache 2.0 (Permissive) 23 GB Gaming GPU
Qwen3.8-27B Alibaba Qwen 27.8B 256K Apache 2.0 (Permissive) 19 GB Gaming GPU
Muse Spark Open Meta 816B 42B active 1M Muse Model License (Conditional) 563 GB GPU server