Ripple-65B Preview in Ollama: How Much Memory the 65B Coding Model Needs

Ollama 0.40.16, released on 9 November 2026, added Ripple-65B Preview, a 65-billion-parameter coding model from Ricursive Intelligence under the Apache 2.0 license. Its smallest Ollama download is 39 GB. That rules out 16 GB and 32 GB computers and any single 24 GB or 32 GB graphics card. A Mac with 64 GB of unified memory, or a PC built around that much fast memory, is the realistic starting point.

What Ricursive released

Ricursive Intelligence builds AI systems for chip design. Ripple-65B Preview is a general-purpose version of the model behind that work, trained for agentic coding: carrying out multi-step programming tasks in a terminal with tools.

Two details need care:

• The context figure differs between pages. Ricursive's model card gives 1,048,576 tokens. Ollama's description says up to 1M tokens, while the table on the same page lists 2M. Treat Ricursive's figure as the model's published limit. The context you can use at home will be far smaller, because conversation memory (KV cache) grows with the context length you set. Context windows explained covers why.
• "Preview" means lightly finished. Ricursive's card says this release has had relatively light post-training, focused on agentic coding, and that its training text was mostly English and code.

Download sizes

Ollama lists these versions. The format and model compression (quantization) decide the size:

The default tag, ripple-preview, is listed at 39 GB to 46 GB: Ollama chooses between the two 4-bit versions, and its page does not say which computers get which file. The 46 GB version is marked MLX, the format Ollama's Apple silicon engine uses. Check the size shown when the download starts, or name the tag you want. GGUF vs MLX explains the difference.

Why the whole 65B counts

Most recent large open models are "mixture of experts" designs, where only part of the model works on each token. Ripple is a dense model: all 65 billion parameters work on every token. Either way, the whole model must be loaded, so the download size is the floor for memory. With a dense model there is no smaller "active" figure to be misled by.

Which computers can hold it

The download is only part of the memory needed. Add conversation memory for your context length, memory for the model runner, and whatever the operating system and your other apps use. The sums below use the published download sizes; Local LLM Guide has not run the model.

Apple silicon Macs (unified memory, shared with macOS and apps):

Windows and Linux PCs:

• One 24 GB or 32 GB graphics card: the 39 GB file is larger than the card's dedicated graphics memory (VRAM). A runner can keep part of the model in system memory (RAM) and run that part on the processor, but answers arrive much more slowly, and you still need roughly 64 GB of RAM.
• Two 24 GB cards (48 GB of VRAM in total): the 39 GB file leaves about 9 GB across both cards for conversation memory, if your runner splits the model between them. That is tight.
• Computers with a large pool of unified memory: Ricursive names the NVIDIA DGX Spark as an example of consumer hardware the model fits on. The new RTX Spark laptops use the same idea; their 128 GB configuration has room, and the 64 GB one is in the same position as a 64 GB Mac.

Not sure what your computer has? What local LLM can my computer run? shows where to find the numbers. The RAM and VRAM calculator lists Ripple-65B Preview; enter your own memory to see whether it fits. The Ripple-65B Preview model guide has the download files and a first task to check.

Run it in Ollama

You need Ollama 0.40.16 or newer. If Ollama is not installed, follow the Ollama installation guide.

ollama --version
ollama run ripple-preview

The second command starts a download of 39 GB to 46 GB. Check your free disk space first. To choose the version yourself, name the tag:

ollama run ripple-preview:65b-q4_K_M

Ollama's page also gives commands that connect the model to coding tools, for example:

ollama launch opencode --model ripple-preview

A coding tool that edits files needs more than a model that chats. Before trusting it with real work, try one small, reversible change in a scratch project, as described in using Continue, Cline or Aider with a local model.

With the MLX version, Ollama says Ripple uses a technique called speculative decoding (a small helper model drafts ahead and the large model checks the draft) to generate text faster. Ollama gives no speed figure for it.

What the benchmark claims cover

These are Ricursive's figures. Local LLM Guide has not tested them.

• Terminal-Bench 2.1: 80.5, which Ricursive presents as the best score among open-weight models from United States developers, ahead of models with seven to fifteen times as many parameters.
• Not first everywhere: in Ricursive's own table, Ripple scores 76.5 on SWE-bench Verified, below three of the four larger models it is compared with.
• Full precision only: Ricursive says every model in its table was tested at full 16-bit precision. That is the 130 GB version. The 4-bit downloads most people will run are compressed, and Ricursive publishes no scores for them.

Ricursive also notes that agentic benchmark scores shifted with the test setup it used. Treat the numbers as a reason to try the model on your own tasks, not as a ranking.

The license

Ripple-65B Preview is released under Apache 2.0. That license permits commercial use, modification and redistribution, provided you keep the license and copyright notices and state significant changes. Converting the model to another format or running it through Ollama does not change those terms.

Keep it local

ollama run ripple-preview runs the model on your own computer after the download. Ricursive also lists the model on hosted services, including Together AI, GMI Cloud, OpenRouter and Fireworks. Prompts sent to those services are processed on their servers. If an app offers "Ripple" without saying where it runs, check before sending private code. The local AI privacy guide explains what to look for.

Also in Ollama 0.40.16

• Qwen3.8 Flash Next uses about 27 GiB less memory on Apple silicon, according to the release notes.
• CUDA 13.5 support for Windows on ARM, which matters for the new RTX Spark computers.
• The OpenAI-compatible /v1/models address returns an empty list, not null, when no models are installed.

Smaller choices

If 39 GB is out of reach, MiMo-V2.6 9B is a coding-focused model with a 6.6 GB download, and choosing a local model for coding covers the sizes in between.

Sources

• Ollama 0.40.16 release notes
• Ollama's Ripple-65B Preview page and its list of versions and sizes
• Ripple-65B Preview model card: design, context, license and benchmark tables
• Ricursive Intelligence's announcement