Ripple-65B Preview

Dense open-weight model for agentic coding and reasoning with 1M context

About model

Ripple-65B Preview has 65B parameters and a 128K-token context window. On Together AI it costs $0.20 per 1M input tokens and $0.60 per 1M output tokens.

Ripple-65B Preview is an open-weight model from Ricursive Intelligence, a general-purpose version of the in-house model powering Ricursive's AI chip design platform. It is a 65B parameter dense Transformer pretrained from scratch on 12 trillion tokens of text and code, with a context window of up to 1M tokens and post-training focused on agentic coding. Ricursive chose a modernized dense architecture over Mixture-of-Experts to reduce training and serving complexity, giving the model strong reasoning performance with low-latency token generation. Ricursive reports a U.S. open-weight state-of-the-art result on Terminal-Bench 2.1, reached with 1.4M reinforcement learning rollouts. Released under Apache 2.0. Available on Together AI.

‍

Dense Parameters

65B

Full capability on every token, built for low-latency generation

Context Window

Long agent trajectories and large codebases in a single session

Open-Weight License

Apache 2.0

Standard license text with no additional terms

Model key capabilities

Agentic Coding: Post-trained with a focus on agentic coding, completing real tasks in terminal environments across long trajectories
Reasoning & Mathematics: Strong performance on competition mathematics and graduate-level science reasoning
Efficient Dense Design: A modernized dense Transformer with sliding-window and global attention, built for strong reasoning with low-latency token generation
Production-Ready Infrastructure: 99.9% SLA, available on serverless and dedicated infrastructure

Performance benchmarks

Model	FrontierMath Tier 4	GPQA Diamond	HLE	SciCode	GDPval-AA	Terminal-Bench 2.1	Agent Arena	FrontierCode	DeepSWE
Ripple-65B Preview		85.2		39.4
Related open-source models
Competitor closed-source models
Claude Fable 5	87.8%	92.6%	53%	60%	62%	85%		53.5%	70%
Claude Opus 5	73.2%	93.2%	53%	56%	68%	89%	+16.4pp	53.4%	74%
GPT-5.6 Sol	82.9%	94.1%	47%	56%	61%	88%		47.5%	73%
Grok 4.5	24.4%	93.1%	40%	54%	51%	82%	+4.2pp	42.4%	54%
GPT-5.6 Luna	61.0%	91.1%	37%	53%	54%	81%		39.8%	67%

API usage

cURL
Python
Typescript

Endpoint:

ricursive-ai/Ripple-65B-Preview

curl -X POST "https://api.together.ai/v1/chat/completions" \
  -H "Authorization: Bearer $TOGETHER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "ricursive-ai/Ripple-65B-Preview",
    "messages": [
      {
        "role": "user",
        "content": "What are some fun things to do in New York?"
      }
    ]
}'

from together import Together

client = Together()

response = client.chat.completions.create(
  model="ricursive-ai/Ripple-65B-Preview",
  messages=[
    {
      "role": "user",
      "content": "What are some fun things to do in New York?"
    }
  ]
)
print(response.choices[0].message.content)

import Together from 'together-ai';
const together = new Together();

const completion = await together.chat.completions.create({
  model: 'ricursive-ai/Ripple-65B-Preview',
  messages: [
    {
      role: 'user',
      content: 'What are some fun things to do in New York?'
     }
  ],
});

console.log(completion.choices[0].message.content);

Model card
Architecture Overview:
• 65B parameter dense Transformer with 80 layers and a model dimension of 5,120
• Sliding-window plus global attention, with gated attention, global NoPE, value residuals, and FFN activation clipping
• Context length of up to 1,048,576 tokens; text input and output, predominantly English
• BF16 weights; knowledge cutoff of December 17, 2025

Training Methodology:
• Pretrained from scratch on 12T tokens of text and code
• Post-training stages: midtraining, long-context training, supervised fine-tuning, long-context SFT, and mixed-domain reinforcement learning, with a focus on agentic coding
• Ricursive reports maintaining 40% MFU through pre-training and post-training with simple sharding and no custom kernels
• Builds on contributions from the Olmo team at the Allen Institute for AI, Prime Intellect, RadixArk, and NVIDIA

Performance Characteristics:
• All figures Ricursive-reported for the instruct model in BF16 at maximum effort
• Agentic coding (Terminal-Bench 2.1, Terminus 2 harness): 80.5, which Ricursive reports as the U.S. open-weight state of the art
• Knowledge and science (MMLU: 89.6; MMLU-Pro: 84.2; GPQA Diamond: 85.2)
• Competition math (AIME 2024: 97.1; AIME 2025: 98.3; AIME 2026: 95.0; HMMT February 2026: 90.5)
• Agentic and applied tasks (τ³-bench banking: 31.3; AutomationBench Public: 36.4; SciCode: 39.4; IFBench instruction following: 79.2; AA-LCR long-context reasoning: 75.0)
• Safety evaluations (FORTRESS Adversarial: 83.8; FORTRESS Benign: 95.5; StrongREJECT: 97.5)
‍
Prompting
Together AI API Access:
• Access Ripple-65B Preview via Together AI APIs using the endpoint ricursive-ai/Ripple-65B-Preview
• Authenticate using your Together AI API key in request headers
• Supports reasoning and tool calling for agentic coding workflows
• Text input and output across a context window of up to 1M tokens
• Available on Together AI serverless and dedicated infrastructure
‍
Applications & use cases
Agentic Coding:
• Run terminal-based coding agents that complete real environment tasks end to end
• Hold large codebases and long trajectories in the 1M-token window
• Chain tool calls across multi-step engineering workflows

Math & Scientific Reasoning:
• Solve competition-level mathematics and graduate-level science problems
• Support scientific coding and analysis tasks
• Serve reasoning workloads where low-latency generation matters

Long-Context Analysis:
• Cross-reference long documents and large working sets in a single request
• Maintain agent state across extended multi-turn sessions
• Process English-language technical material at scale
‍