Ripple-65B Preview: Local Inference, Agentic Coding

1. A General-Purpose Model Focused on Coding

Ripple-65B Preview is a general-purpose model. Its initial release underwent relatively light post-training focused on agentic coding, where the model uses tools to carry out coding tasks.

The model has 65 billion parameters, was pretrained on 12 trillion tokens, and supports a context of up to 1 million tokens.

The release includes an instruction-tuned model and a base model that has only undergone pretraining. The instruction-tuned model comes in high-precision BF16 and compact, low-precision NVFP4 formats. The company says it provides NVFP4 checkpoints, files containing saved model states, and Colab resources for local development. The license is Apache 2.0.

Distribution channels include Hugging Face and cloud services. The company says the model is available through Together AI, GMI Cloud, and OpenRouter.

2. Leading Coding Scores Under a Shared Evaluation Harness

The company reports a Terminal-Bench 2.1 score of 80.52. The model card's table shows 80.5 and says comparison models were evaluated at BF16 precision with the model's reasoning effort set to maximum.

The model card lists company-reported scores and results using a shared evaluation harness in separate rows. The comparison below uses Terminus 2 for task execution and scoring, with a 128K sequence length and summarization.

Table: Coding benchmark using a shared evaluation harness

Conditions: Terminal-Bench 2.1 (Terminus 2) row in the company's model card, BF16 precision, maximum effort, official Terminus 2 harness, 128K sequence length, summarization enabled

Unit: Score

In this company-reported comparison, Ripple scored 1.8 points above runner-up Muse Spark Open at 78.7. The model card says scores vary with the harness and summarization settings, so adoption decisions should account for the listed evaluation conditions.

3. Trading Simpler Architecture for Per-Token Compute

To simplify training and serving infrastructure, the company chose a dense architecture that uses all parameters per token rather than a mixture of experts that uses only some. Serving means operating a model to respond to requests. Internally, the design uses narrower layers stacked more deeply.

This choice trades the cost of processing 65 billion parameters per token for a total model size that fits on local hardware and a simpler operating structure. The mixture-of-experts models compared by the company are larger overall but activate 23 billion to 55 billion parameters per token. Total size indicates whether a model fits in device memory, while active parameters per token indicate compute demands during generation.

The company reports total R&D costs of $32 million, from development through final training. That is one-tenth of its estimated cost for the previous leading open-weight model for agentic coding. It says it did not buy post-training data, instead building its own synthetic data pipeline to automate data generation and processing.

4. Running on Local Devices and in Data Centers

The company says the model fits on consumer hardware such as NVIDIA DGX Spark. It describes DGX Spark as having more compute relative to memory bandwidth than data center accelerators. Ripple is designed to use that compute while reducing total storage requirements; being dense alone does not make it better suited to local execution. In a data center, the company reports 250 tokens per second on two B200 GPUs with a batch size of 8, the number of inputs processed together. This was before quantization, which reduces storage by lowering numerical precision, and custom kernels were applied.

The server guide provides an example pairing the BF16 model with a DFlash2 draft model. The draft model proposes candidates for upcoming tokens in advance.

The runtime uses SGLang 0.5.19 and requires Linux, CUDA Toolkit 13, a C++ compiler, and uv. The code below is reproduced exactly from the repository's server guide.

export CUDA_HOME=/usr/local/cuda-13.0
export PATH="$CUDA_HOME/bin:$PATH"

CUDA_VISIBLE_DEVICES=0,1 uv run --script serve.py \
  --model-path /path/to/target \
  --speculative-draft-model-path /path/to/draft

5. Developers Can Start by Choosing a Deployment Format

For local development, start with the low-precision checkpoints and Colab resources. For a data center server, start with the two-GPU setup guide. The practical change in this release is a path from model selection to self-hosting, supported by both public weights and runnable examples.

Ripple-65B Preview connects competitive coding performance with local deployment while requiring users to consider both model storage size and runtime compute demands.