A Guth Labs publication ↗
Back to News

Ripple-65B Preview gives local AI builders an Apache-licensed coding model

AI-generated editorial illustration of cyan ripples radiating across copper circuitry on a dark silicon slab.
AI-generated editorial illustration by guthlabs.newsroom

Ricursive Intelligence released Ripple-65B Preview on November 4, putting a new coding-focused model into the hands of developers who want to run and adapt their own AI. The company’s dated launch announcement identifies the release as a general-purpose version of the model behind its chip-design platform.

The release combines a dense 65-billion-parameter Transformer with downloadable weights. Ricursive says it trained the model from scratch on 12 trillion tokens and supports a context window of up to 1,048,576 tokens. Its model card lists instruction-tuned BF16 and NVFP4 checkpoints alongside a BF16 base model.

Downloadable weights with a permissive license

The repository includes an Apache 2.0 license, which permits use, modification and redistribution subject to its conditions. There is no non-commercial restriction in that license. Redistribution still carries obligations, including preserving applicable notices and identifying modified files.

That gives builders a concrete option to evaluate on infrastructure they control. The distinction matters for teams choosing between a hosted model and one they can customize, pin to a particular version and keep serving independently.

Local deployment still requires a memory budget

The BF16 instruction model is substantial. Its Hugging Face file listing totals about 130 GB and contains 16 safetensors shards. Download size is not a complete runtime-memory estimate: an inference deployment also needs room for context, caches and its serving software.

Ricursive’s smaller NVFP4 checkpoint is explicitly marked experimental. It quantizes the feed-forward layers while retaining attention in BF16. Builders should therefore avoid treating the checkpoint as a uniformly four-bit model, or assuming its name alone determines the memory required.

Ricursive points to NVIDIA DGX Spark as a local deployment target in its launch post. That is a useful starting point for evaluation, rather than a promise that the model will fit every developer workstation.

Coding results are promising and vendor-reported

Ricursive reports 80.52 on TerminalBench 2.1 and positions Ripple against much larger US open-weight models. The company also reports $32 million in total research and development costs, including recipe development and the final training runs, rather than describing that figure as the cost of a single pretraining run. Both figures come from Ricursive’s own account.

The comparison needs care. The model card’s methodology combines published scores with internal evaluations and says harness choices can move results. Its strongest coding result should not be read as proof that Ripple leads every workload, or that a quantized local installation will reproduce a BF16 benchmark.

Start with the current inference code

The release is already receiving practical fixes. The repository history records Transformers inference changes on November 6 and generation, masking and cache-related fixes on November 10. Those changes make the exact revision part of any reproducible evaluation.

For a useful first trial, builders should pin the checkpoint and runtime, measure memory at their intended context length, and compare task completion, latency and failures on their own repositories. Ripple’s immediate value is the ability to conduct that test with accessible weights and a permissive license. Its place in a production stack will depend on the results.

Sources and citations

Revision history

Published .