Ricursive's Ripple-65B Preview is a 65B open model that beats far bigger US rivals at agentic coding

Ricursive reports a TerminalBench 2.1 score of 80.52, ahead of much larger US open-weight models including Reflection Beam 501B (80.1), Meta Muse Spark Open 816B (76.4), Thinking Machines Inkling 975B (63.8) and NVIDIA Nemotron 3 Ultra 550B (56.4)[1]. These are vendor-run results and haven't been independently verified yet[4].

The model is a general-purpose variant of the proprietary model that powers Ricursive's AI chip design platform[1]. The company says the Preview has had "relatively light posttraining, with a focus on agentic coding"[2].

Rather than a mixture-of-experts design, Ricursive chose a modernised dense Transformer to keep training and serving simple[2]. It uses a thin-and-deep shape, with a narrow 5,120-wide hidden size and 80 layers, plus gated attention, activation clipping and value residuals[2]. Starting from an Olmo 3 baseline, the team says it improved compute efficiency on code by 16x across its model ladder, and it held about 40% hardware utilisation with no custom kernels while growing from one to five people[2]. Neural Brew reports that Ricursive puts the model's total research and development cost at $32 million[4].

Four repositories went up on Hugging Face: the BF16 instruct model at about 130 GB, an experimental NVFP4 build at about 57 GB, a pretrain-only base model, and a 4B draft model for speculative decoding[3]. Ricursive says the model fits on an NVIDIA DGX Spark and reaches 250 tokens a second on two B200 GPUs before quantisation[2]. It's also hosted on Together AI, GMI Cloud and OpenRouter, with Fireworks to follow[1]. One catch worth knowing: OrcaRouter notes the first commercial endpoint advertises a 16,384-token context, far short of the 1M the weights support[3].

Most of the strongest US open-weight models are now hundreds of billions of parameters. A 65B dense model that competes with them on agentic coding, under a permissive license and small enough for a desktop AI box, gives teams a realistic option for running a capable coding agent on their own hardware.