Ripple-65B, released by Ricursive Intelligence on 4 November 2026, scores 80.52 on TerminalBench 2.1, just ahead of Reflection Beam 501B at 80.1, according to Ricursive's announcement. Those are company-reported figures, but if they hold, a 65B dense model matches a rival roughly eight times its size. Our verdict: for local agentic coding experiments on modest hardware, Ripple-65B is now the strongest US open-weight option to try first; for repo-scale work, larger closed models are still the safer choice.
Last verified: 5 November 2026.
TL;DR
- Ripple-65B Preview is a 65B-parameter dense Transformer with Apache 2.0 weights, published on Hugging Face.
- Ricursive reports 80.52 on TerminalBench 2.1, against 80.1 for Reflection Beam 501B (Ricursive).
- Hosted pricing on OpenRouter is $0.20 per 1M input tokens and $0.60 per 1M output tokens, about ₹18.9 and ₹56.6 at ₹94.3 to the dollar (Investing.com).
- The 1M-token context is a company claim; the OpenRouter endpoint serves 16K today.
- No independent leaderboard has scored the model yet, so treat the benchmark lead as provisional.
How does Ripple-65B compare with larger open-weight models?
On Ricursive's own TerminalBench 2.1 runs, Ripple-65B sits first among the open-weight models it chose to compare, despite being the smallest. The table below uses the company's published comparison.
| Model | Parameters | TerminalBench 2.1 |
|---|---|---|
| Ripple-65B Preview | 65B | 80.52 |
| Reflection Beam | 501B | 80.1 |
| Meta Muse Spark Open | 816B | 76.4 |
| NVIDIA Nemotron 3 Ultra | 550B | 56.4 |
| Thinking Machines Inkling | 975B | 63.8 |
Source: Ricursive announcement, company-reported. The model card notes that some competitor scores come from Ricursive's own in-house runs, which is worth remembering before reading the gaps as settled.
TerminalBench 2.1 itself is an independent benchmark: 89 curated tasks run in real terminal environments, maintained by Stanford and Laude Institute researchers (Terminal-Bench), and scored as per-task pass rates. On the public leaderboard, frontier agent-model pairs such as GPT-5.3-Codex with Codex CLI reach 79.1 (Terminal-Bench), so Ricursive's 80.52 claim, if it replicates, would place this 65B open model in frontier-agent territory. The useful reading is the shape of the table rather than the decimal places. Ripple and Reflection Beam are effectively tied, and the two largest models, Muse Spark Open and Inkling, are more than ten times Ripple's size yet score lower. Parameter count is a weak predictor of agentic coding results here.
Why does a 65B dense model score this well?
Ricursive attributes the result to training efficiency rather than scale. According to the announcement, the model was pretrained on 12 trillion tokens, and reached its 80.5 score with 1.4M reinforcement-learning rollouts, which the company says is roughly 70 times fewer than Reflection Beam needed.
The architecture is plain by design. It is a dense Transformer, not a mixture-of-experts, with a thin-and-deep layout of d_model 5120 and 80 layers. Ricursive says it trained in BF16 at around 40% model FLOPs utilisation without custom kernels, and that the model is a general-purpose variant of the one behind its AI chip-design platform (Ricursive).
What does the Ripple-65B price look like?
The headline figure is the training bill: Ricursive says the full R&D cost was $32 million, covering pretraining, posttraining, data and evaluations, and describes it as one-tenth of the previous open-weight leader's cost (Ricursive). At ₹94.3 to the dollar that is about ₹302 crore (Investing.com). Spread across 65,000 million parameters, it works out at roughly $492 per million parameters, our own arithmetic from the company's figure.
For users, the relevant numbers are the API rates on OpenRouter:
- Input: $0.20 per 1M tokens, about ₹18.9
- Output: $0.60 per 1M tokens, about ₹56.6
- Cached input: $0.05 per 1M tokens, about ₹4.7
Rupee figures use the 5 November 2026 close of 94.347 per dollar (Investing.com); the IMF's representative rate for 3 November was 94.8606 (IMF), so expect small drift. OpenRouter lists no launch discount, and GMI Cloud was the only provider on the listing at the time of our check. Ricursive also names Together AI and Fireworks as hosts.
Can you run Ripple-65B locally?
Yes, and this is the strongest argument for it. The weights are public and ungated under Apache 2.0 on Hugging Face. The repository includes the BF16 instruction-tuned model, an NVFP4 quantised checkpoint with attention kept in BF16, a 4B DFlash2 drafter for speculative decoding, and the pretrain-only Base Preview, public since 14 October 2026.
Ricursive says the NVFP4 build fits NVIDIA DGX Spark class hardware and provides a Colab notebook to get started. For speed, the company quotes 250 tokens per second on two B200 GPUs at batch size 8 before quantisation (Ricursive). The hosted OpenRouter endpoint shows 89 tokens per second at P50 with 0.56 seconds P50 latency (OpenRouter). Local speed on a DGX Spark will differ from both, and we have not measured it.
Where does Ripple-65B fall short?
Three limits matter before you build on it.
- Unverified scores. No independent leaderboard has evaluated it, so the TerminalBench lead rests on Ricursive's own runs (ArtificialWatch).
- Context. The 1M-token window is the company's claim about the model. The OpenRouter listing serves 16K, which is tight for agents that read whole repositories (OpenRouter).
- Preview status. There is no announced date for a non-Preview release.
Ricursive does not claim to beat the largest closed models on research-level maths or repository-scale engineering, and our own view is that the Preview label should be taken at face value.
Which should you pick?
- Local testing of agentic coding on consumer or workstation hardware: Ripple-65B, using the NVFP4 checkpoint.
- Cheap hosted experiments with short contexts: the OpenRouter endpoint, within its 16K limit.
- Long-context or production coding agents: stay with established options until independent scores arrive. Our comparisons in Qwen vs Claude vs Kimi and Kimi K3.1 Pro vs Claude Opus 5.5 cover the stronger alternatives.
For the wider tooling picture, see best AI agent for coding in 2026 and agentic AI coding tools in 2026.
FAQ
Q: What is Ripple-65B?
A: Ripple-65B is a 65B-parameter dense Transformer from Ricursive Intelligence, released as a Preview on 4 November 2026 with Apache 2.0 weights on Hugging Face (Ricursive).
Q: How much does Ripple-65B cost to use?
A: On OpenRouter it costs $0.20 per 1M input tokens and $0.60 per 1M output tokens, about ₹18.9 and ₹56.6 at current rates (OpenRouter).
Q: Does Ripple-65B really beat Reflection Beam 501B?
A: On Ricursive's own TerminalBench 2.1 runs it scores 80.52 against 80.1, which is a narrow lead and not yet independently verified (Ricursive).
Q: Can I run Ripple-65B on my own hardware?
A: Ricursive says its NVFP4 checkpoint fits NVIDIA DGX Spark class hardware, and a Colab notebook is provided for getting started (Ricursive).
Q: Is the 1M-token context available?
A: Not on OpenRouter, where the listing serves 16K tokens; the 1M figure is the company's claim about the model (OpenRouter).
Corrections log
No corrections to date. Rupee conversions use the 5 November 2026 exchange rate and will drift.
This article was produced with AI assistance and edited against primary sources. See our editorial process on the How We Work page.
Top comments (0)