Ricursive Intelligence, a company founded by former Google DeepMind researchers Anna Goldie and Azalia Mirhoseini, has released the Ripple-65B Preview model for agentic coding in open preview. The dense transformer, featuring 65 billion parameters without Mixture of Experts (MoE), was trained on 12 trillion tokens with a context window of up to 1 million tokens. It scored 80.52 on Terminal-Bench 2.1, outperforming larger open MoE models by an order of magnitude. Weights are available under the Apache 2.0 license on Hugging Face, and APIs are accessible via Together AI, GMI Cloud, and Fireworks at 20/60 cents per million tokens. The company estimates the development cost at approximately $32 million. On SWE-bench Verified and GPQA Diamond, the model currently trails Reflection Beam, and all published figures remain vendor-reported, awaiting independent verification.

image
image
image

What Happened

Ricursive Intelligence opened the preview for its agentic coding model, Ripple-65B Preview. This is a dense transformer with 65 billion parameters and no mixture of experts: 80 layers, d_model 5120, pre-training on 12 trillion tokens, and a context window of up to 1 million tokens. The model scored 80.52 on Terminal-Bench 2.1, outperforming Reflection Beam (501 billion parameters, 80.1), Muse Spark (816 billion, 76.4), and Thinking Machines' Inkling (975 billion, 63.8). The company claims this result is the SOTA among US open-weight models for agentic coding. However, on SWE-bench Verified, the model scored 76.5 compared to Reflection Beam's 81.6, and on GPQA Diamond, it scored 85.2 compared to 90.6. Ricursive Intelligence estimates the development cost at approximately $32 million, which is roughly 10 times cheaper than the previous open SOTA. Weights are published under Apache 2.0 on Hugging Face, and APIs are available through Together AI, GMI Cloud, and Fireworks at 20/60 cents per million tokens.

Context

Ricursive Intelligence founders Anna Goldie and Azalia Mirhoseini previously worked on the AlphaChip project at Google DeepMind, making the focus on computational efficiency a logical continuation of their experience. Efficiency is built directly into the architecture: an increased proportion of layers with sliding window attention, a redundant number of attention heads, activation limits in FFN blocks, gated attention, and a complete abandonment of positional encoding. According to the company, on the dolma_stack dataset, cross-entropy improves by up to 16 times compared to standard pre-training. The data recipe is also non-standard: the proportion of code in pre-training is significantly increased, and post-training was performed entirely on a proprietary synthetic pipeline with quality and coverage metric filtering, without using purchased data. Behavioral refinement was conducted via reinforcement learning on 1.4 million rollouts, which is approximately 70 times less than Reflection Beam. This combination of decisions explains why a mid-sized dense model is being compared to giant-scale MoE models.

Why This Matters for the Industry

The main signal for the industry is not the benchmark victory, but the economics. A model built roughly an order of magnitude cheaper than the previous open SOTA, with an RL phase compact by tens of times compared to the nearest open competitor, has reached the level of giant-scale open MoE models in agentic coding. If independent runs confirm these figures, the economics of open weights will shift: front-end agentic coding will no longer require MoE scale and hundreds of millions of dollars, and the combination of code-centric data, synthetic post-training, and modest RL will become a template for new open models. A possible scenario in tooling is that agentic frameworks, IDE plugins, and routers will establish the 65B class as a cheap executor layer paired with a large MoE model for complex cases. Finally, such a low API price sets a new lower threshold for the unit economics of agentic products and can accelerate the overall price reduction for agentic inference.

Why This Matters for Users

The model can be tried today via three paths. Weights under Apache 2.0 can be downloaded from Hugging Face and run locally: the model fits on consumer hardware like the NVIDIA DGX Spark, and on 2xB200 it reaches 250 tokens per second. If you don't have your own hardware, it is easier to connect to APIs from Together AI, Fireworks, or GMI Cloud: 20 cents per million input tokens and 60 cents per million output tokens make piloting cheap. With this, it is realistic to build a prototype in your own codebase in one working day—a CLI hook for commits, a review bot for pull requests, or auto-closing tickets with terminal access. Before integrating the model into production, it is advisable to run it and your current working model on your repositories and compare the percentage of successfully completed tasks, cost per task, and latency: Ripple-65B Preview's strength is specifically terminal agentic cycles, while its general scores drop, so the decision should be based on your measurements, not vendor benchmarks.

What Is Still Unknown / Limitations

All figures on Terminal-Bench 2.1, SWE-bench Verified, and GPQA Diamond are currently vendor-reported: independent runs on the released weights have not been published. The 0.42-point lead on Terminal-Bench 2.1 is claimed without confidence intervals or task breakdowns, meaning statistically it is a noisy margin rather than a solid victory, and the thesis that agentic coding no longer requires MoE scale remains a hypothesis until independent replications. The development cost estimate, pre-training efficiency, and RL phase scale are also company claims: a full technical report that would allow them to be verified has not yet been released. The quality of the synthetic post-training pipeline and the absence of data contamination from benchmarks have not yet been independently verified. Finally, this is a preview version of the model, not a final release, so its characteristics may differ from the full version.

Sources

Author

Look at AI, editorial team