Ricursive releases Ripple-65B Preview

Ricursive’s X post announcing Ripple-65B Preview was timestamped Nov. 4, 2026.[2] The company describes the release as a general-purpose variant of the proprietary model that powers its AI chip design platform.[1] Ricursive says the model is dense, has 65 billion parameters, and was pretrained from scratch on 12 trillion tokens.[2][1] It supports a context length of up to one million tokens.[1] These are company descriptions of the model; the announcement does not establish how it will perform on every workload.

Ricursive reports a TerminalBench 2.1 score of 80.52, higher than the scores it lists for Reflection Beam 501B, Muse Spark Open 816B, Inkling 975B, and Nemotron 3 Ultra 550B.[1] The company also reports scores of 79.1 on SWE-Bench Pro v2.0-Hard and 44.3% on DeepSWE v1.1.[2] It says it reached a TerminalBench score of 80.5 with 1.4 million rollouts, about 70 times fewer than Reflection Beam.[1] An outside review said the benchmark results had not yet been checked by an independent lab.[4] Treat these figures as reported results, not independent confirmation.

Ricursive says the full research-and-development effort cost $32 million, covering work from scratch, development, and final training runs; it estimates that this was one-tenth the cost of the previous open-weight agentic-coding leader.[1] The company says it did not buy post-training data and instead built synthetic-data pipelines.[1] It also reports 40% model-flop utilization during pretraining and post-training without custom kernels or quantization.[1] Ricursive says the dense design aims to combine strong reasoning with low-latency token generation.[1] Those cost and efficiency figures are company claims.

Ricursive lists Hugging Face, Together AI, GMI Cloud, and OpenRouter as places where the Preview is available, and said Fireworks would go live with it later that day.[1] Eesel reported hosted pricing of $0.20 per million input tokens, $0.60 per million output tokens, and $0.05 per million cached input tokens.[4] Ricursive says the model can run efficiently on local hardware such as NVIDIA DGX Spark.[1] On two B200 GPUs at batch size eight, the team reports generation speed of 250 tokens per second before quantization or custom kernels.[3] The company says the team grew from one person to five.[3]