Key takeaway
Ricursive Intelligence released Ripple-65B Preview on November 6, 2026: a 65B dense Transformer pretrained on 12 trillion tokens with up to 1 million tokens of context, scoring 80.52 on TerminalBench 2.1, which the company calls the best result for a US open-weight model on that benchmark, at a full from-scratch R&D cost of $32 million.
What happened
On November 6, 2026, Ricursive Intelligence published Ripple-65B Preview on its official site. The company builds AI for chip design and says it trained a language model from scratch so its chip design platform could handle agentic coding, then released a general-purpose variant.
The headline claim is US open-weight state of the art: Ripple-65B Preview scores 80.52 on TerminalBench 2.1, above larger models including 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 model is available through Hugging Face, Together AI, GMI Cloud, OpenRouter, and Fireworks, and the company says it runs efficiently on local hardware such as NVIDIA DGX Spark.
Key details
| Item | What Ricursive published |
|---|---|
| Release date | November 6, 2026 |
| Parameters | 65 billion, dense Transformer rather than a mixture of experts |
| Pretraining data | 12 trillion tokens |
| Context length | Up to 1,000,000 tokens |
| TerminalBench 2.1 | 80.52 |
| Comparison scores | Reflection Beam 501B 80.1, Meta Muse Spark Open 816B 76.4, Inkling 975B 63.8, Nemotron 3 Ultra 550B 56.4 |
| R&D cost | $32 million, covering ladder experiments, data mixing, synthetic data, post-training, and the final pretraining and post-training runs |
| RL scale | 1.4 million rollouts, about 1/70 of Reflection Beam according to the company |
| Availability | Hugging Face, Together AI, GMI Cloud, OpenRouter, Fireworks |
Architecture and training
Ripple-65B uses a dense architecture instead of a mixture of experts. Ricursive's stated reason is lower infrastructure complexity for pretraining and post-training: it sustained 40% model FLOPs utilization (MFU) throughout, with no custom kernels and no quantization.
The published configuration is a thin-and-deep shape: d_model of 5120 and 80 layers. To improve reasoning, the model overprovisions attention heads while raising the ratio of sliding-window attention (SWA) layers to global layers and shrinking the SWA window to offset the cost. It also adds FFN activation clipping, gated attention, global NoPE, and value residuals.
The pretraining recipe came from a series of model-ladder experiments. Ricursive says it started from the Allen Institute for AI's Olmo 3 ladder and added StepFun batch-size and learning-rate scaling laws, smooth dimension and depth scaling, pre+post norm, sigmoid gated attention, global NoPE, and value residuals, then switched to a custom dataset with substantially upsampled code. The compounded effect: at 4B scale, matching the code performance of the Olmo 3 ladder took up to 16 times less compute, measured by dolma_stack cross-entropy.
For post-training, Ricursive says it bought no post-training data and instead 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 that get harder as the model improves. The company reports reaching 80.5 on TerminalBench 2.1 from its initial SFT checkpoint with only 1.4 million rollouts.
Cost and comparisons
The $32 million figure drew the most attention. Ricursive defines it as full R&D cost from scratch, including pretraining ladder development, data mixing, evals, debugging, synthetic data generation, post-training recipe development, and the final hero pretraining and post-training runs. The company says this is about one-tenth the cost of the previous open-weight state of the art on agentic coding, citing public reporting that Reflection committed $150 million per month for compute at SpaceX Colossus.
Note the sourcing: the 80.1 for Reflection Beam 501B and 76.4 for Meta Muse Spark Open 816B come from Ricursive's own comparison table, so they are a vendor-compiled cross-comparison, and independent reproductions are not public. TerminalBench 2.1 is a public benchmark, but harness settings, tool permissions, and prompts may differ across labs, so cross-vendor scores are not a clean capability ranking.
On parameter efficiency, a 65B dense model scoring above 501B and 816B open-weight models is the company's central claim. Its explanation has two parts: the architecture and training recipe were tuned for coding, and RL needed roughly 1/70 of the rollouts, which suggests data and task design mattered more than raw compute.
Impact and limits
For users, the immediate change is deployability. A 65B dense model can run on a single machine or a small cluster, and Ricursive specifically cites low-memory-bandwidth devices such as NVIDIA DGX Spark, in contrast with 500B-plus models that need multi-GPU clusters.
For the open-weight ecosystem, the release fills a "medium-parameter, coding-specialized" slot. The high end of open-weight leaderboards is dominated by MoE models above 400B, so a 65B dense model in the same score range gives teams that need local or private coding agents another option.
Facts and assessments should be separated: 80.52, $32 million, and 1.4 million rollouts all come from Ricursive's own post and are vendor-reported. "US open-weight state of the art" refers to a single leaderboard, TerminalBench 2.1, not an aggregate ranking across benchmarks. The company also did not publish complete results for long-context retrieval, multilingual performance, or safety evaluations.
Official sources
- Ricursive Intelligence: Introducing Ripple-65B Preview, a Frontier Open-Weight Model (November 6, 2026)