Glonce

Ricursive releases Ripple-65B Preview, claims top U.S. open-weight score on TerminalBench 2.1

modelsThe Glonce editorsHacker News🔥79Glonce rating79Impact65Novelty62Relevance100Surprise55Players35Credibility75

Ricursive Intelligence, which describes itself as a full-stack AI lab building self-improving systems by co-optimizing hardware, software and model, has released Ripple-65B Preview. The company says it trained the language model from scratch to power its AI chip design platform. The release is a general-purpose variant of the proprietary model behind that platform, and an internal variant still runs the platform.

The model was pretrained on 12 trillion tokens and supports up to 1M context length. Ricursive says it scores 80.52 on TerminalBench 2.1, a new U.S. open-weight state of the art on agentic coding. It lists four larger U.S. open-weight models as comparisons: 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). The lead over Reflection Beam is 0.42 points on Ricursive's own numbers.

The headline claim is cost. Ricursive notes that Sam Altman said in 2023 that GPT-4 training exceeded $100 million, and that Dario Amodei later said frontier models can cost up to $1 billion to train, with pretraining costs projected at $10-100B by 2026 to 2027. Against that, Ricursive says it developed and trained Ripple-65B Preview for only $32M. That figure covers everything from scratch: pretraining ladder development, data mixing, evals, debugging, synthetic data generation, posttraining recipe development and the final pretraining and posttraining run. The company says this is one-tenth the cost of the previous open-weight SOTA on agentic coding, "as estimated to the best of our knowledge." A footnote says U.S. open-weight players like Reflection were reported to commit $150M per month for compute at SpaceX Colossus, and that Reflection reported Beam took 4 weeks to pretrain and 4 weeks to posttrain.

On architecture, Ricursive chose a modernized dense Transformer rather than Mixture-of-Experts. It says this kept training simple: 40% MFU throughout pretraining and posttraining, with simple sharding, no custom kernels and no quantization. The model is thin and deep, with d_model 5120 and 80 layers. To improve reasoning, it overprovisions attention heads, raises the ratio of sliding-window to global attention layers and shrinks the sliding window to offset the cost. It also adds FFN activation clipping, gated attention, global NoPE and value residuals. The dense design gives low-latency token generation, and the small parameter envelope lets the model run on local devices such as NVIDIA DGX Spark.

On the post-training side, Ricursive says it bought no posttraining data and built its own synthetic data pipelines, quality filtering and coverage metrics. The mixed RL recipe uses truncated importance sampling, score centering and a linear length penalty, with long-horizon synthetic environments whose tasks got harder as the model improved. It says it reached 80.5 on TerminalBench 2.1 with only 1.4M rollouts, roughly 70x fewer than Reflection Beam.

Pretraining was tuned through a series of small-scale "ladder" experiments that started from Olmo 3's ladder. Steps included StepFun batch size and learning rate scaling laws, a ladder that smoothly scales architecture dimensions and depth, pre+post norm, sigmoid gated attention, global NoPE with a final LR cooldown to zero, value residuals, and a custom dataset with substantially upsampled code. One idea, 2-stream mHC, looked good at small scale but did not help at 4B and was rejected. Together the changes improved compute efficiency on code, measured by dolma_stack CE, by up to 16x at 4B scale; the company says the final ladder reaches similar code performance to Olmo 3's with 16x less compute. It also argues the gain should grow with scale, because the Olmo 3 baseline slope is shallower.

Ripple-65B Preview is available on Hugging Face, Together AI, GMI Cloud and OpenRouter, with Fireworks going live later the same day. Ricursive credits open-source work from the Olmo team at AI2, Prime Intellect, RadixArk and NVIDIA, and says it is hiring researchers and engineers.

Key facts

  • Ricursive Intelligence released Ripple-65B Preview, a dense (not MoE) open-weight model pretrained on 12 trillion tokens with up to 1M context.
  • The company reports 80.52 on TerminalBench 2.1, a U.S. open-weight best, ahead of 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).
  • Ricursive says full R&D cost was $32M, which it estimates is one-tenth the cost of the previous open-weight SOTA on agentic coding.
  • It says RL reached 80.5 with 1.4M rollouts, roughly 70x fewer than Reflection Beam, and that training held 40% MFU with no custom kernels or quantization.
  • Weights are on Hugging Face, Together AI, GMI Cloud and OpenRouter, with Fireworks due later the same day.

Why it matters

The claim is that a smaller model can match significantly larger U.S. open-weight models on agentic coding at a fraction of the cost. Ricursive puts Ripple-65B Preview at 80.52 on TerminalBench 2.1 against 80.1 for Reflection Beam 501B, while its own model is far smaller than every comparison it lists (the others range from 501B to 975B in their names). It frames the $32M full R&D cost against public estimates of $100 million to $1 billion and more for frontier training. It also bets on a dense Transformer rather than Mixture-of-Experts, arguing that simplicity kept training efficient.

Who it affects

Developers building coding agents who want an open-weight model they can call through hosted providers or run locally. Ricursive says the model's size suits devices such as NVIDIA DGX Spark. It also touches other U.S. open-weight labs, since Ricursive names Reflection, Meta, Thinking Machines and NVIDIA as comparisons, and teams who train their own models and may look at the ladder and RL recipe details.

How to use it

Ripple-65B Preview is available on Hugging Face, Together AI, GMI Cloud and OpenRouter, and Fireworks goes live later the same day. The model supports up to 1M context length. Ricursive says it can run efficiently on local devices such as NVIDIA DGX Spark. The source does not say whether the weights licence is permissive or give the licence name, so check the Hugging Face page before building on it.

How solid is it

This is a first-party announcement. The benchmark scores, including the competitors' numbers, and the $32M and one-tenth cost figures are all Ricursive's own claims; no independent verification is mentioned. The one-tenth comparison is hedged by the company itself as an estimate "to the best of our knowledge," and the source gives no breakdown of the $32M. The competitor cost comparison rests on reported figures for Reflection's compute commitment and training time in a footnote. The release is labelled a Preview, and only TerminalBench 2.1 results are given.

Risks and caveats

A single benchmark carries the headline, and the margin over Reflection Beam is 0.42 points on Ricursive's numbers. The source does not say how the competitor scores were obtained. The 16x compute-efficiency gain is a maximum, measured by dolma_stack CE at 4B scale against Olmo 3's ladder, not a result on the final model, and the claim that it grows with scale is Ricursive's inference. The source does not say how many or which GPUs were used for training. The licence is not stated. The $32M covers R&D as the company defines it, so it may not be comparable to other labs' quoted costs.

“we developed and trained Ripple-65B Preview for only $32M, which is one-tenth the cost of the previous open-weight SOTA on agentic coding, as estimated to the best of our knowledge.”

— Ricursive Intelligence, release post