Release of Ripple-65B Preview

Ricursive Intelligence has announced the release of its first large language model (LLM), a 65-billion parameter dense model titled “Ripple-65B Preview.” According to technical documentation and community discussions on platforms such as Reddit and Hugging Face, the model is a pre-training-only checkpoint that has undergone relatively light post-training specifically focused on agentic coding tasks.

The model was trained on 12 trillion tokens and supports a context length of up to one million tokens. Technical specifications indicate a “thin and deep” architecture consisting of 80 layers and 5,120 dimensions, utilizing features such as gated attention, global NoPE, and value residuals.

Performance Benchmarks

The release positions the 65B model as a significant competitor in the open-weight sector, particularly regarding agentic coding capabilities. Unverified reports from community discussions suggest that Ripple-65B may outperform several much larger models in specific tasks, including Muse Spark Open-816B, Beam-501B, Inkling-975B, and Nemotron 3 Ultra-550B.

Reported benchmark results for the Ripple-65B Preview include:

  • TerminalBench 2.1: 80.5%
  • SWE-Bench Pro v2.0-Hard: 79.1% (though some community sources cite a lower figure of 61.4%)
  • DeepSWE v1.1: 44.3%

When compared to existing models like Qwen 3.8-27B, the Ripple-65B shows higher performance in Terminal-Bench 2.1 (80.5% vs 73.0%) but trails in GPQA Diamond (85.2% vs 89.2%) and IFBench (79.2% vs 79.5%).

Technical Implementation and Accessibility

The model is currently available via Hugging Face. Early adopters have noted that because the model is a dense architecture, it may require significant hardware resources to run efficiently. Community members have discussed the practicalities of local deployment, noting that a Q8_0 quantization requires approximately 68.8 GB of memory, while a Q4_K_M quantization requires 38.8 GB.

While the model is being released as an open-weight preview, some observers have noted the high cost of development, with unverified claims suggesting approximately $35 million in venture capital funding was utilized for its creation.

The release has prompted technical debate regarding the model’s utility in its current state. Because the checkpoint is a “pretrain-only” version without extensive synthetic QA data or midtraining, some users have questioned its immediate conversational utility, while others suggest that its current performance in coding tasks may only represent a fraction of its long-term potential once full post-training is applied.