65B (dense)
Up to 1,048,576 tokens; shipped model configuration sets 2,097,152 positions
Apache License 2.0
Open-weight, text-only general-purpose preview derived from the proprietary model powering Ricursive Intelligence's AI chip-design platform. Ripple-65B Preview is a dense Transformer pretrained from scratch on 12 trillion predominantly English and code tokens, followed by 200B mid-training tokens, dedicated long-context training, 183B supervised-fine-tuning tokens, long-context supervised fine tuning, and mixed-domain reinforcement learning. Its thin-and-deep architecture uses 80 layers, a 5,120-dimensional hidden state, 40,960-dimensional feed-forward layers, 80 attention heads, eight key/value heads, a 1,024-token sliding-attention window, and a 100,352-token vocabulary. Ricursive describes the design as Gemma-3n-like and adds pre-plus-post normalization, FFN activation clipping, sigmoid-gated attention, global NoPE layers, value residuals, an increased sliding-window-to-global-attention ratio, and a reduced sliding-window size. The model is advertised for up to one million tokens of context, while the shipped configuration encodes 2,097,152 positions and the tokenizer defaults to 131,072 tokens. Ricursive reports that its final model ladder improves code compute efficiency by 16 times at 4B scale and by up to 55 times for the full run relative to its starting OLMo 3 ladder. The company says the team grew from one to five people, sustained approximately 40% model FLOP utilization using simple sharding without custom kernels, and spent $32 million on total research and development including the final training runs. Post-training used synthetic data generated internally and a mixed reinforcement-learning recipe with truncated importance sampling, score centering, linear length penalties, long-horizon environments, and approximately 1.4 million rollouts. Publisher-reported results include 80.52 on Terminal-Bench 2.1, 76.5 on SWE-bench Verified, 61.42 on SWE-bench Pro v1.0, 79.1 on SWE-bench Pro v2.0 Hard, 44.3 on DeepSWE 1.1, 71.67 on MCP Atlas, 79.2 on IFBench, and 75.0 on AA-LCR. Ricursive cautions that reproducing modern agentic benchmarks is difficult and reports internal measurements when official comparison scores are unavailable. The official releases comprise a 129.566 GB BF16 instruction checkpoint across sixteen safetensor shards, a 129.566 GB pretrain-only BF16 base checkpoint, a 57.214 GB NVFP4 instruction checkpoint, and an 8.076 GB DFlash2 speculative-decoding model. The supplied SGLang server supports separated reasoning content, structured tool calls, retained reasoning history, and BF16 target-plus-draft serving on two NVIDIA B200 GPUs. It does not expose a thinking toggle or multiple reasoning-effort levels. Ricursive reports approximately 250 tokens per second at batch size eight on two B200 GPUs before quantization or custom-kernel optimization and positions the smaller dense parameter envelope for local systems such as NVIDIA DGX Spark.