Ricursive Ripple-65B hits 80.5 on Terminal-Bench 2.1
Ricursive Intelligence has released Ripple-65B Preview, a 65 billion parameter dense open-weight language model. The company says the model is pretrained on 12 trillion tokens and supports up to 1 million context length. It scores 80.52 on Terminal-Bench 2.1, which Ricursive calls U.S. open-weight state of the art on agentic coding. The model is a general-purpose variant of the one that powers Ricursive's chip design platform. It is available on Hugging Face, Together AI, GMI Cloud, and OpenRouter. Fireworks goes live later, Ricursive says.
This matters because agentic coding models are usually large. Ripple-65B Preview beats several bigger open models. On Terminal-Bench 2.1, Reflection Beam 501B scores 80.1, Meta Muse Spark Open 816B scores 76.4, Thinking Machines Inkling 975B scores 63.8, and NVIDIA Nemotron 3 Ultra 550B scores 56.4. Ripple's 80.52 beats all of them with 65 billion parameters compared to Reflection Beam's 501 billion. Ricursive says the dense architecture gives strong reasoning with low latency token generation, and the small parameter envelope lets it run efficiently on local devices like NVIDIA DGX Spark.
In one line: Ricursive released a 65B dense open-weight model that scores 80.52 on Terminal-Bench 2.1, beating larger US open models on agentic coding and costing $32 million to train.
Why the score matters
Terminal-Bench 2.1 tests agentic coding, which is more than generating a function. It asks the model to work on longer, multi-step tasks. A high score here says the model can hold context and make decisions while coding, not just complete a snippet. Ripple-65B Preview gets 80.52, which is slightly better than Reflection Beam's 80.1 but with a fraction of the parameters. The model also runs as a dense Transformer, not a mixture of experts, so every inference uses all 65 billion parameters. That can mean more predictable latency and easier deployment. For teams that run coding agents locally, this is interesting. If you keep a list of coding tools, the Coding and developer tools directory is a good place to check what else exists.
Training for less than one-tenth the cost
Ricursive says its full research and development cost was $32 million. That includes pretraining ladder development, data mixing, evals, debugging, synthetic data generation, posttraining recipe development, and the final hero pretraining and posttraining run. The company estimates this is one-tenth the cost of the previous open-weight state of the art on agentic coding. The source notes that Reflection was reported to commit to $150 million per month for compute at SpaceX Colossus, and that Reflection Beam took four weeks to pretrain and four weeks to posttrain.
Ricursive did not buy any posttraining data. It built its own synthetic data pipelines, quality filtering, and coverage metrics. The RL recipe uses truncated importance sampling, score centering, and a linear length penalty with long-horizon synthetic environments. Ricursive says it reached a Terminal-Bench 2.1 score of 80.5 with only 1.4 million rollouts, roughly 70 times fewer rollouts than Reflection Beam. For a look at how agentic coding can change client delivery work, see A brief goes in, software comes out.
Architecture choices that helped
Ricursive chose a modernized dense Transformer instead of a mixture of experts. The model is thin and deep, with a narrow d_model of 5120 and 80 total layers. To improve reasoning, Ricursive overprovisioned the number of attention heads while increasing the sliding window attention to global layer ratio and reducing the sliding window size. It added FFN activation clipping, gated attention, global NoPE, and value residuals. These choices kept model FLOPs utilisation at 40 percent throughout pretraining and posttraining, even with simple sharding, no custom kernels, and no quantization.
The team also optimized its pretraining recipe through model laddering. It started from Olmo 3's ladder and tested interventions at small scale. The final version improved compute efficiency on code as measured by dolma_stack cross entropy by up to 16 times at 4 billion parameter scale. In other words, Ricursive says the final ladder can achieve similar code performance to the Olmo 3 ladder with 16 times less compute. That is a meaningful gain for anyone trying to train or fine-tune smaller coding models.
What it does not yet do
This is a preview release. The source does not include an independent benchmark suite beyond Terminal-Bench 2.1, so we cannot yet judge how the model handles general language, long context, or safety. Ricursive does not detail licence terms in the provided material, which matters for commercial use. The model supports up to 1 million context length, but no long context evaluation is reported here. It is also tuned for agentic coding, so it may not be the best choice for every task. Treat the 80.52 as a strong result on one benchmark, not a full audit.
Practical take
For client RAG and agentic coding work, I would watch Ripple-65B Preview for two things. First, whether the license allows commercial use and fine-tuning. Second, whether independent teams reproduce the Terminal-Bench 2.1 score and test it on real codebases. The $32 million training cost and 40 percent MFU show what can be done with careful engineering, and that may push more teams to try local coding models. I would test it on a small internal coding task first, not on client work. Nothing to do yet, just watch the license and the reproductions.
Quick answers
What is Ricursive Ripple-65B Preview?
Ricursive Ripple-65B Preview is a 65 billion parameter dense open-weight language model pretrained on 12 trillion tokens with up to 1 million context. It scores 80.52 on Terminal-Bench 2.1, which Ricursive says is U.S. open-weight state of the art for agentic coding.
How much did Ricursive spend to train Ripple-65B Preview?
Ricursive says its full R&D cost was $32 million, including pretraining ladder development, data mixing, evals, synthetic data, posttraining, and final runs. The company estimates this is one-tenth the cost of the previous open-weight state of the art on agentic coding.
Where can I download or run Ripple-65B Preview?
Ripple-65B Preview is available on Hugging Face, Together AI, GMI Cloud, and OpenRouter. Fireworks goes live later, Ricursive says. The model is designed to run efficiently on local devices like NVIDIA DGX Spark.