Ricursive trains 65B parameter model for $32M with new coding benchmarks

Ricursive Intelligence has released Ripple-65B Preview, a new open-weight language model designed to power its AI chip design platform. The company trained this 65-billion parameter model from scratch for $32 million, achieving a new United States open-weight state-of-the-art score on agentic coding benchmarks while significantly undercutting the training costs of larger competitors.

What happened

Ricursive Intelligence, a full-stack AI lab focused on co-optimizing hardware and software, announced the public release of Ripple-65B Preview. This general-purpose variant is derived from the proprietary model that internally powers their AI chip design tools. The model was pretrained on 12 trillion tokens and supports context lengths of up to one million tokens. Upon release, it immediately set a new benchmark for U.S. open-weight models in agentic coding tasks.

The performance metrics place Ripple-65B ahead of several significantly larger models. It achieved a score of 80.52 on TerminalBench 2.1, surpassing 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). Despite having fewer parameters than these competitors, Ripple-65B delivers strong reasoning performance with low-latency token generation due to its dense architecture. The model is currently available on Hugging Face, Together AI, GMI Cloud, and OpenRouter, with Fireworks support scheduled to go live shortly after the announcement.

Perhaps most notable for the industry is the cost efficiency. While estimates for frontier model training have ranged from $100 million to over $1 billion, Ricursive developed and trained Ripple-65B Preview for just $32 million. This figure represents the total research and development cost, including pretraining ladder development, data mixing, evaluations, debugging, synthetic data generation, and the final training runs. The company states this is approximately one-tenth the cost of the previous open-weight state-of-the-art model for agentic coding.

How it works

The efficiency gains stem from a combination of architectural choices and a disciplined training strategy known as "model laddering." Rather than adopting Mixture-of-Experts (MoE) architectures common in recent large models, Ricursive opted for a modernized dense Transformer. This decision reduced infrastructure complexity and allowed them to maintain a solid Model FLOPs Utilization (MFU) of 40% throughout training without requiring custom kernels or quantization. The architecture features a "thin-and-deep" aspect ratio, with a narrow hidden dimension of 5120 and 80 total layers. To boost reasoning capabilities, they overprovisioned attention heads, increased the ratio of Sliding Window Attention (SWA) to global layers, and reduced the SWA window size. Additional technical modifications included FFN activation clipping, gated attention, global NoPE (No Positional Embeddings), and value residuals.

Training efficiency was further enhanced through a steep ascent reinforcement learning (RL) approach. Instead of relying on brute-force RL scaling or purchasing post-training data, the team developed internal synthetic data pipelines with rigorous quality filtering. They employed a mixed RL recipe using truncated importance sampling, score centering, and linear length penalties within high-quality, long-horizon synthetic environments. As the model improved, harder tasks were continuously introduced. This method allowed Ripple-65B to reach its top TerminalBench score with only 1.4 million rollouts, which is roughly 70 times fewer than the prior state-of-the-art model, Reflection Beam.

The pretraining recipe was refined through iterative small-scale experiments called model ladders. Starting from the Olmo 3 model ladder, the team tested various interventions, such as switching from post-norm to pre-plus-post norm, implementing sigmoid gated attention, and adding value residuals. These compounded changes improved compute efficiency on code by up to 16x at the 4-billion parameter scale compared to the baseline Olmo 3 ladder. The final version utilized a custom pretraining dataset with substantially upsampled code, ensuring the model specialized effectively in coding tasks before scaling up.

Key details

• Model Size: 65 billion parameters, dense architecture.
• Training Cost: $32 million total R&D spend, including data generation and infrastructure.
• Benchmark Score: 80.52 on TerminalBench 2.1, a new U.S. open-weight state-of-the-art.
• Context Length: Supports up to 1 million tokens.
• Training Data: Pretrained on 12 trillion tokens with a focus on upsampled code.
• Efficiency: Achieved SOTA performance with 70x fewer RL rollouts than the previous leader.

Why it matters

For engineering teams and technical founders, Ripple-65B demonstrates that massive parameter counts are not strictly necessary for high-performance agentic coding. The success of a dense, thin-and-deep architecture challenges the prevailing trend toward sparse MoE models, suggesting that careful architectural tuning can yield better latency and efficiency on specific hardware. This is particularly relevant for developers looking to run capable models on local devices or edge hardware, such as the NVIDIA DGX Spark, where memory bandwidth is often a bottleneck compared to raw compute FLOPs.

The cost breakdown also provides a realistic benchmark for what is achievable with disciplined engineering. By avoiding expensive data purchases and focusing on synthetic data quality and efficient RL recipes, Ricursive achieved frontier-level performance at a fraction of the projected industry costs. This signals a shift where algorithmic efficiency and data quality may outweigh brute-force compute scaling, offering a more sustainable path for organizations building specialized AI tools without access to hundred-million-dollar budgets.

What you can do

• Download Ripple-65B Preview from Hugging Face to test its agentic coding capabilities in your local environment.
• Evaluate the model on your own coding benchmarks to compare its latency and accuracy against larger MoE models you currently use.
• Review the architectural details, such as the thin-and-deep ratio and global NoPE implementation, to inform your own model design choices.
• Experiment with the model on hardware with lower memory bandwidth, such as local workstations, to assess its deployment feasibility for edge applications.
• Study the reported RL recipe, specifically the use of synthetic data and truncated importance sampling, to optimize your own post-training pipelines.
• Monitor the upcoming availability on Fireworks to integrate the model into cloud-based inference workflows if local hosting is not viable.