OnPrem World地端世界

Models Published:

Ricursive releases Ripple-65B Preview, a dense Apache 2.0 coding model sized for a DGX Spark, now in Ollama

Illustration: Ricursive releases Ripple-65B Preview, a dense Apache 2.0 coding model sized for a DGX Spark, now in Ollama
Illustration

What happened

Ricursive Intelligence released Ripple-65B Preview on November 4, a 65B dense Transformer pretrained from scratch on 12 trillion tokens with up to 1,048,576 tokens of context, published on Hugging Face under the Apache 2.0 licence. It is a general-purpose version of the model behind Ricursive's AI chip-design platform, released as an instruction-tuned model in BF16 and NVFP4 plus a BF16 base model. Ricursive reports 80.52 on Terminal-Bench 2.1, ahead of the much larger open models in its own comparison table, and says the whole project cost US$32 million. Ollama v0.40.16 added it on November 9 as `ollama run ripple-preview`, with a 39 GB Q4_K_M build and a 46 GB NVFP4 build for its MLX engine.

Why it matters locally: A strong agentic coding model at 65B fits in the memory of a single desktop AI box, and its permissive Apache 2.0 licence allows commercial use on a company's own code.

Three ways to read it

For the market

Ricursive compares Ripple only with US open-weight models, and claims to beat ones with up to 975B total parameters on Terminal-Bench 2.1; many of the figures come from labs' own reports or Ricursive's in-house runs, so independent results will decide whether that holds. If it does, it would be a notable sign that a small team with a dense model can compete with the large MoE releases that dominate open weights today.

In plain words

Ripple is an AI helper for programmers that can carry out multi-step jobs in a terminal on its own. It is small enough to run on one desktop AI computer in the office, so the code it works on never has to go to an outside service.

For practitioners

Unlike MoE coding models, all 65B parameters are active on every token, so generation speed on memory-bandwidth-limited machines such as the DGX Spark (273 GB/s in our data) will likely trail MoE models of similar size; on Ollama's MLX engine, Ripple uses DFlash 2 speculative decoding to generate faster. The 39 GB Q4_K_M or 46 GB NVFP4 builds leave room for a long context on a 64 GB or 128 GB machine, while the BF16 weights need about 130 GB. Test it on your own repositories and harness before trusting the self-reported scores.

Sprout's insight

Sprout compares this with other brands and with where real applications are heading. Free for members.

Sign in free to read it →

Subscribe to Sprout's weekly

One email a week: the top stories and a taste of Sprout’s insights. Unsubscribe any time.

Source

Read the official source: Introducing Ripple-65B Preview (Ricursive Intelligence) ↗

Summaries are written by OnPrem World from the official source. We link to the original and never republish it.

On this site

More news

All news →