Ricursive Releases Ripple-65B Preview

Ricursive Intelligence released Ripple-65B Preview, a general-purpose open-weight language model, saying it achieved U.S. open-weight state of the art on agentic coding. The company says its proprietary model powers its AI chip design platform and that it trained Ripple from scratch for $32 million in full research and development costs. The model was pretrained on 12 trillion tokens and supports a context length of up to 1 million tokens. On TerminalBench 2.1, Ricursive reports a score of 80.52, ahead of Reflection Beam 501B at 80.1, Meta Muse Spark Open 816B at 76.4, Thinking Machines Inkling 975B at 63.8, and NVIDIA Nemotron 3 Ultra 550B at 56.4.

Ricursive says the $32 million covered development of its pretraining ladder, data mixing, evaluations, debugging, synthetic data generation, post-training methods, and final training runs. It calls this one-tenth the cost of the previous open-weight agentic-coding leader, based on its estimate. As background cost comparisons, the company cites Sam Altman’s 2023 statement that GPT-4 training exceeded $100 million and Dario Amodei’s following-year estimate that frontier model training could reach $1 billion, with pretraining costs projected at $10 billion to $100 billion by 2026 to 2027. A footnote says Reflection was reported to commit $150 million per month for compute at SpaceX Colossus and reported that Beam took 4 weeks each to pretrain and post-train.

For reinforcement learning, Ricursive says it developed synthetic data pipelines and did not buy post-training data. Its recipe combined truncated importance sampling, score centering, and a linear length penalty with long-horizon synthetic environments; it adjusted task difficulty as the model improved. The company reports reaching 80.5 on TerminalBench 2.1 after 1.4 million rollouts, about 70 times fewer than Reflection Beam.

Ricursive designed the model for training and serving on NVIDIA hardware, choosing a dense Transformer rather than a mixture-of-experts architecture. It reports 40% model FLOP utilization during pretraining and post-training with simple sharding, no custom kernels, and no quantization. The model has 65 billion parameters, 80 layers, and a 5,120-wide d_model; architectural choices also include more attention heads, a higher sliding-window-attention to global-layer ratio, a smaller sliding window, FFN activation clipping, gated attention, global NoPE, and value residuals. Ripple is available through Hugging Face, Together AI, GMI Cloud, and OpenRouter; Fireworks was scheduled to go live later that day.

The company says experiments based on Olmo 3’s model ladder tested successive changes, including scaling rules, architecture and depth scaling, pre- and post-normalization, sigmoid gated attention, 2-stream mHC, global NoPE, a final learning-rate cooldown, value residuals, and a final dataset with more code. It rejected 2-stream mHC after it failed to improve results at 4 billion parameters. Ricursive says the combined changes improved compute efficiency on code, measured by dolma_stack cross-entropy, by up to 16 times at 4 billion parameters, allowing similar code performance with 16 times less compute than the Olmo 3 ladder. The company says an internal variant powers its chip-design platform and credits open-source contributions from AI2, Prime Intellect, RadixArk, and NVIDIA.

Ricursive Intelligence released Ripple-65B Preview, a general-purpose open-weight language model, saying it achieved U.S. open-weight state of the art on agentic coding. The company says its proprietary model powers its AI chip design platform and that it trained Ripple from scratch for $32 million in full research and development costs. The model was pretrained on 12 trillion tokens and supports a context length of up to 1 million tokens. On TerminalBench 2.1, Ricursive reports a score of 80.52, ahead of Reflection Beam 501B at 80.1, Meta Muse Spark Open 816B at 76.4, Thinking Machines Inkling 975B at 63.8, and NVIDIA Nemotron 3 Ultra 550B at 56.4.

Ricursive says the $32 million covered development of its pretraining ladder, data mixing, evaluations, debugging, synthetic data generation, post-training methods, and final training runs. It calls this one-tenth the cost of the previous open-weight agentic-coding leader, based on its estimate. As background cost comparisons, the company cites Sam Altman’s 2023 statement that GPT-4 training exceeded $100 million and Dario Amodei’s following-year estimate that frontier model training could reach $1 billion, with pretraining costs projected at $10 billion to $100 billion by 2026 to 2027. A footnote says Reflection was reported to commit $150 million per month for compute at SpaceX Colossus and reported that Beam took 4 weeks each to pretrain and post-train.

For reinforcement learning, Ricursive says it developed synthetic data pipelines and did not buy post-training data. Its recipe combined truncated importance sampling, score centering, and a linear length penalty with long-horizon synthetic environments; it adjusted task difficulty as the model improved. The company reports reaching 80.5 on TerminalBench 2.1 after 1.4 million rollouts, about 70 times fewer than Reflection Beam.

Ricursive designed the model for training and serving on NVIDIA hardware, choosing a dense Transformer rather than a mixture-of-experts architecture. It reports 40% model FLOP utilization during pretraining and post-training with simple sharding, no custom kernels, and no quantization. The model has 65 billion parameters, 80 layers, and a 5,120-wide d_model; architectural choices also include more attention heads, a higher sliding-window-attention to global-layer ratio, a smaller sliding window, FFN activation clipping, gated attention, global NoPE, and value residuals. Ripple is available through Hugging Face, Together AI, GMI Cloud, and OpenRouter; Fireworks was scheduled to go live later that day.

The company says experiments based on Olmo 3’s model ladder tested successive changes, including scaling rules, architecture and depth scaling, pre- and post-normalization, sigmoid gated attention, 2-stream mHC, global NoPE, a final learning-rate cooldown, value residuals, and a final dataset with more code. It rejected 2-stream mHC after it failed to improve results at 4 billion parameters. Ricursive says the combined changes improved compute efficiency on code, measured by dolma_stack cross-entropy, by up to 16 times at 4 billion parameters, allowing similar code performance with 16 times less compute than the Olmo 3 ladder. The company says an internal variant powers its chip-design platform and credits open-source contributions from AI2, Prime Intellect, RadixArk, and NVIDIA.