Ripple-65B Preview: 65B Apache-2.0 Agentic Coding Model With Weights Already Live
Ricursive Intelligence released Ripple-65B Preview, a dense 65B open-weight coding model that reportedly outperforms far larger models on agentic coding benchmarks.
Chip designer Ricursive Intelligence has released Ripple-65B Preview, a dense 65B transformer trained on 12 trillion tokens with a context window of up to 1 million tokens. The model ships under a full Apache-2.0 license with no usage clauses, and the BF16, NVFP4, and base checkpoints are already downloadable on Hugging Face, alongside community GGUF quantizations that appeared within 48 hours of launch.
The vendor reports a TerminalBench 2.1 score of 80.52, ahead of several far larger open-weight models, plus 79.1% on SWE-Bench Pro v2.0 Hard and 71.67% on MCP Atlas. The company claims a total R&D cost of about $32 million and says the model runs at 250 tokens per second on a two-GPU B200 setup without quantization. Third-party serving is available through Together AI, OpenRouter, Fireworks, and GMI at $0.20 per million input tokens and $0.60 per million output tokens.
The release is notable because it challenges the assumption that competitive agentic coding requires a large GPU cluster: a 65B dense model fits on a single mid-range server. The results remain vendor-reported, and press coverage notes the model trails Inkling 2 Small on several coding benchmarks, so independent evaluation is still needed.