Ricursive AI Ripple-65B Tops Open-Source Coding Benchmark

Ricursive Intelligence put out Ripple-65B Preview on Thursday, a 65-billion-parameter dense model trained from scratch for agentic coding work. The company says it scored 80.52 on Terminal-Bench 2.1, the highest self-reported mark among open-weight models built in the US. That number carries weight because of what it's beating: every Mixture-of-Experts rival ahead of it on that benchmark runs between 501 billion and 975 billion total parameters. Ripple-65B does it with a fraction of that footprint.

The release ships three checkpoints under Apache 2.0, an instruction-tuned BF16 model, an NVFP4 build aimed at NVIDIA DGX Spark hardware, and a BF16 base model. Ricursive trained it on 12 trillion tokens with context length reaching up to 1 million, and the model runs at 250 tokens per second on two B200 GPUs. It's a general-purpose offshoot of the proprietary model that powers Ricursive's AI chip design platform, which gives some indication of where the company's engineering priorities actually sit.

The bigger story might be what it cost to build, and why Ricursive picked a dense architecture when nearly everyone else chasing this kind of performance went with MoE.