Ripple-65B Preview: Ricursive's $32M Open-Weight Model Claims U.S. SOTA on Agentic Coding
Ricursive Intelligence released Ripple-65B Preview, a 65B-parameter dense open-weight language model that the company says achieves U.S. open-weight state of the art on agentic coding, scoring 80.52 on TerminalBench 2.1. The headline claim is not just the benchmark number but the price: Ricursive says it developed and trained the model from scratch for $32M in total R&D cost, roughly one-tenth the cost of the previous open-weight SOTA on agentic coding, to the best of its knowledge.
Ripple-65B Preview is a general-purpose variant of the proprietary model that powers Ricursive's AI chip design platform. Ricursive describes itself as a full-stack AI lab building self-improving systems by co-optimizing hardware, software, and model.
What Ripple-65B Preview is
Ripple-65B Preview is a dense Transformer with 65B parameters, pretrained on 12 trillion tokens with support for up to 1M context length. Ricursive optimized the architecture for its own training and serving platforms rather than adopting a Mixture-of-Experts design.
The model is available on Hugging Face, Together AI, GMI Cloud, and OpenRouter, with Fireworks going live later on the release day.
The benchmark claim: U.S. open-weight SOTA on TerminalBench 2.1
Ricursive reports a TerminalBench 2.1 score of 80.52 for Ripple-65B Preview and compares it against four larger U.S. open-weight models:
| Model | Parameters | TerminalBench 2.1 |
|---|---|---|
| Ripple-65B Preview | 65B (dense) | 80.52 |
| Reflection Beam | 501B | 80.1 |
| Meta Muse Spark Open | 816B | 76.4 |
| Thinking Machines Inkling | 975B | 63.8 |
| NVIDIA Nemotron 3 Ultra | 550B | 56.4 |
The comparison is explicitly scoped to U.S. open-weight models. Ricursive does not include non-U.S. open-weight models in the chart, a choice that drew substantial criticism in the Hacker News discussion.
Training cost: $32M for the full R&D run
Ricursive states that $32M covers its full R&D cost starting from scratch, including pretraining ladder development, data mixing, evals, debugging, synthetic data generation, posttraining recipe development, and the final pretraining and posttraining runs.
The company frames this against historical estimates for frontier training: Sam Altman's 2023 statement that GPT-4 training exceeded $100 million, and Dario Amodei's projection that pretraining costs could reach $10–100B by 2026–2027. Ricursive's footnote points to reporting that Reflection committed $150M per month for compute at SpaceX Colossus, and to Reflection's own report that Beam took four weeks to pretrain and four weeks to posttrain.
The RL recipe: 1.4M rollouts and a 70x reduction
Ricursive says it deliberately avoided a brute-force reinforcement learning climb and instead iterated on SFT and RL recipes for compute efficiency. The company did not buy any posttraining data, instead building 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, run against high-quality long-horizon synthetic environments. Tasks were tuned to the model's capabilities, with harder tasks added continuously as the model improved.
Ricursive reports reaching a TerminalBench 2.1 score of 80.5 with only 1.4M rollouts, which it describes as roughly 70x fewer rollouts than Reflection Beam, the prior SOTA.
Architecture: dense, thin-and-deep, and built for simple infrastructure
Ripple-65B Preview uses a modernized dense Transformer rather than a Mixture-of-Experts architecture. Ricursive says the goal was to reduce pretraining and posttraining infrastructure complexity while preserving reasoning and agentic performance.
Key architectural choices reported by the company:
- Thin-and-deep aspect ratio with a narrow
d_modelof 5120 and 80 total layers. - Overprovisioned attention heads, with a higher sliding-window-attention-to-global layer ratio and a reduced SWA window size to offset the added cost.
- FFN activation clipping, gated attention, global NoPE, and value residuals.
- 40% MFU sustained through pretraining and posttraining, achieved with simple sharding, no custom kernels, and no quantization.
Model laddering: 16x compute efficiency on code versus the Olmo 3 ladder
Ricursive optimized its pretraining recipe through a sequence of model laddering experiments, starting from the Olmo 3 model ladder and testing interventions at small scale with a focus on coding.
The ladder progressed through StepFun batch size and learning rate scaling laws (v1.0), a new ladder that smoothly scales architectural dimensions and depth (v1.1), a switch from post-norm to pre+post norm that was neutral at small scale but improved from 4B (v1.2), sigmoid gated attention (v1.3), a 2-stream mHC variant that was rejected because it did not improve at 4B scale (v1.4), global NoPE plus a final LR cooldown to zero (v1.5), and value residuals (v1.6). The final version switched to a custom pretraining dataset with substantially upsampled code.
Ricursive says the compounded changes improved compute efficiency on code, as measured by dolma_stack cross-entropy, by up to 16x at 4B model scale — meaning the final ladder reaches similar code performance to the Olmo 3 ladder with 16x less compute. The company notes that the Olmo 3 baseline slope is shallower than its own ladder, so the efficiency gain should increase with scale.
What the Hacker News discussion challenged
The dominant critique was the narrow definition of "SOTA." Commenters argued that scoping the claim to U.S. open-weight models makes the benchmark less useful for end users comparing against the real field.
"Category narrowing to this degree really has diluted the term 'SOTA' quite a lot. You mean it's state of the art for this subset of benchmarks you're targeting? No? Well surely it's state of the art for models of a certain size category? Huh, what about models that are available for local inference? O, state of the art for open models of a certain size made in the US specifically. Uh. Neat!" — @advael
"I understand the desire to look the best in some self defined category. But not showing any comparison against the actual leaders in open weight renders all of the benchmarking completely information-less for the actual end users (who care about real performance comparison against the real field)." — @a11r
Several commenters specifically flagged the absence of Qwen comparisons. One supplied numbers the post omitted: Qwen 3.8 Flash Next scores 84.3 on TerminalBench 2.1, and Qwen 3.8-27B scores 73.0.
"This model scores 80.5 on Terminal Bench 2.1. For reference Qwen 3.8 Flash Next (which is bigger than this model but probably cheaper to run due to MoE) scores 84.3, and Qwen 3.8-27B scores 73.0. I think this model could be competitive, depending on how much it was benchmaxxed compared to Qwen?" — @2001zhaozhao
"When a post doesn't mention Qwen for comparison, you know the performance isn't going to be great." — @syntaxing
Others questioned the benchmark selection itself. One commenter argued TerminalBench 2.1 is old, SWE-bench was disavowed by OpenAI, and DeepSWE is benchmark-hacked by Gemini. Another called the comparison chart "befuddling" and accused the post of cherry-picking weaker models.
A counterpoint defended the U.S.-only framing on compliance grounds:
"There are several situations where some of us are explicitly prohibited from using models from companies not under the jurisdiction of the United States. It doesn't mean these models are competitive when looked at from a global view but it is also not an irrelevant differentiation." — @bluejay2387
Local inference: dense 65B is a different tradeoff than MoE
Commenters noted that a 65B dense model is expensive to run on consumer hardware compared with MoE models of similar or larger total size, because all parameters are active per token. One commenter speculated the target buyer is an organization doing chip design in air-gapped environments that can afford server-grade inference hardware.
"As is, it's probably hard to become popular amongst hobbyists and local model enthusiasts like Qwen's 27B class models or MOE models because it's a 65B dense model and inference will be prohibitively slow on consumer grade hardware... this is for the companies who want to use AI in air-gapped environments for chip design so that they protect their IPs." — @stewartpark
A contrasting report came from a high-memory Mac Studio user, who said the model's size is actually a good fit for Mac clusters:
"In my experience, the sweet spot for Mac Clusters is around 200GB of weights, which this is far under. Congratulations to the recursive team for building an open model that will actually run!" — @thesnarkitecht
Cost, not capability, may be the real story
One commenter argued the model is unremarkable on raw performance but that the training economics are the notable result, extrapolating that a model of this class could cost under $5M within three years — potentially within reach of crowd-funded efforts.
Open questions
- How much of the 80.52 is benchmark-specific tuning? Ricursive describes RL tasks tuned to its model's capabilities and harder tasks added as the model improved, which is standard practice but makes cross-model comparison harder without independent evaluation.
- How does Ripple-65B Preview compare to non-U.S. open-weight leaders? The release does not say, and commenters supplied Qwen numbers that place at least one Qwen model above Ripple's reported score.
- Does the dense architecture hold up at larger scale? Ricursive argues its ladder's steeper slope means efficiency gains grow with scale, but the 16x figure is measured at 4B model scale.
- What does the chip-design variant do? Ricursive says an internal variant powers its AI chip design platform but has not published benchmarks for that variant.
Ricursive credits the Olmo team at AI2, Prime Intellect, RadixArk, and NVIDIA as open-source foundations for the work, and says it is hiring researchers and engineers to push "full-stack AI from model to silicon."