Ripple-65B Preview
Ripple-65B Preview is a 65B-parameter open-weight model from Ricursive Intelligence for agentic coding, reasoning and tool calls. On Wiro a request takes about 15K input tokens and returns up to 8K.
Overview
Ripple-65B Preview is a 65B-parameter dense Transformer from Ricursive Intelligence, released as open weights under Apache 2.0. It was pretrained on about 12 trillion tokens of mostly English text and code, then post-trained with a focus on agentic coding. Ricursive reports 80.52 on TerminalBench 2.1 for it. Wiro runs the instruction-tuned BF16 checkpoint.
What you can build
- Coding agents that plan a change, call your tools and continue from the results
- Code review and refactoring plans for files and modules of up to about 15K tokens
- Technical Q&A that benefits from working through the problem before answering
- Assistants that look up data through your functions, then answer from it
- Summaries and cross-checks of documents of up to about 15K tokens
Inputs
- Your request as plain text, or a conversation as a list of messages.
- Optional system instructions that set rules and tone.
- Optional user and session identifiers that keep chat history between turns.
- Optional function definitions, a tool choice setting and the results of earlier tool calls.
- Sampling controls: randomness, nucleus threshold, top-k, repetition control and length bias.
- Maximum and minimum output length in tokens, with stop strings and a numeric seed.
- A toggle that switches between deterministic decoding and sampling.
Outputs
The model returns text for the assistant message. When it reasons first, the reasoning comes back as a separate section before the final answer. For a tool turn it returns the function name and its JSON arguments. It calls one function per turn, so a task that needs several tools runs over several turns.
Recommended settings
- Leave the maximum output length at 0 to let the budget follow your prompt.
- For repeatable output, turn sampling off or set a seed.
Limitations
- On Wiro, a single request takes in about 15,000 input tokens, counting the system message, chat history, tool definitions and your prompt together. The checkpoint's larger window is not available here.
- Longer input is shortened automatically, without an error: the oldest history messages are dropped first, then the end of your latest message is cut. Put your question at the start of long prompts.
- Plan for about 8,000 output tokens per request. Reasoning counts toward them.
- The model takes text only. It does not read images, audio or video.
- Its training data is mostly English and code, so other languages can be weaker.
- This is a preview release with relatively light post-training.
- Stop strings can cut off answers mid-sentence if they match common text.
Safety & compliance
The weights are released under Apache 2.0. Check generated code before running it, and validate tool arguments before your app acts on them. Don't send secrets or private personal data unless your policy allows it.
API quick start
Run Ripple-65B-Preview with a single API call.
{
"prompt": "Write a Python function that merges overl…",
"user_id": "...",
"session_id": "...",
"system_prompt": "You are a helpful, respectful and honest …"
}curl -X POST "https://api.wiro.ai/v1/Run/ricursiveai/ripple-65b-preview" \
-H "Content-Type: application/json" \
-H "x-api-key: YOUR_WIRO_API_KEY" \
--data-binary @- <<'JSON'
{
"prompt": "Write a Python function that merges overl…",
"user_id": "...",
"session_id": "...",
"system_prompt": "You are a helpful, respectful and honest …"
}
JSONOther models you might like
Explore similar llm & chat models.
