Disclosure: this article contains no affiliate links. All links are plain links, and we earn nothing if you click them. See our affiliate disclosure. Based on Ricursive’s official launch page, Hugging Face model card and OpenRouter listing, checked November 14, 2026.
AI-assisted: AI helped gather and draft this research; the facts were checked against the cited sources.
Quick answer: Ripple-65B Preview is an open-weight model from Ricursive Intelligence, aimed at agentic coding. The weights went up on Hugging Face on November 4, 2026 under the Apache 2.0 licence, so they’re free to download. If you’d rather not host it yourself, OpenRouter lists it from $0.10 per million input tokens and $0.30 per million output tokens on its cheapest provider, with up to a 1,048,576-token context window on two of its three providers. Ricursive’s own page says the model supports up to 1M tokens, but that’s a vendor claim.
What Ricursive released
According to Ricursive’s launch page, Ripple-65B Preview is “a general-purpose variant” of the private model behind its AI chip design platform. The company says:
- it trained the model from scratch on 12 trillion tokens
- it’s a dense Transformer, not a Mixture-of-Experts model
- it supports “up to 1M context length”
- the model is on Hugging Face, Together AI, GMI Cloud, OpenRouter and Fireworks
Ricursive also says the whole project, from first experiments to the final training run, cost $32 million. That’s Ricursive’s own figure. We haven’t seen anyone else check it.
The Hugging Face repository carries the apache-2.0 licence tag and isn’t gated, so you don’t need to request access. The release includes an instruction-tuned model in BF16 and NVFP4, plus a BF16 base model.
About the benchmark numbers
Ricursive reports a score of 80.52 on TerminalBench 2.1 and says that beats several larger US open-weight models — 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). These numbers are vendor-reported and haven’t been independently verified. We found no outside test results, so we’re not using them to rank Ripple against anything.
Ways to try it
- Download the weights for free. The Hugging Face repo has them under Apache 2.0, which allows commercial use. You’ll need serious hardware. Ricursive’s own example of a local device is the NVIDIA DGX Spark, not a normal laptop (launch page).
- Pay per use on OpenRouter. The OpenRouter listing shows $0.10 per million input tokens and $0.30 per million output tokens on its cheapest provider (Prime Intellect), with Together and GMICloud at $0.20 and $0.60. Tool calling works on two of the three providers. We didn’t see a free API tier.

- Other hosts. Ricursive says Together AI, GMI Cloud and Fireworks also host it. We haven’t checked their prices, so none are quoted here.
What a test run costs
Using OpenRouter’s cheapest rates, say you send 100 prompts with 2,000 input tokens and 500 output tokens each. That’s 200,000 input tokens ($0.02) plus 50,000 output tokens ($0.015), about $0.035 in total.
The catches
- Context length. Ricursive claims up to 1M tokens. On OpenRouter, two of the three providers now offer 1,048,576 tokens and Together offers 131,072, so the 16,384-token limit we saw on November 5 is gone.
- The benchmarks are vendor-reported only. Run it on your own tasks before you rely on it.
- Self-hosting is a big job. Unless you already have the hardware, running a 65-billion-parameter model yourself isn’t realistic.
- The name. The company is “Ricursive”, not “Recursive”. Check you’re on the right site or repo.
Who it’s for
Building small coding or automation agents? Ripple is a cheap model to add to your tests. There’s no consumer chat app, though, so if you just want a ready-made tool, start with something like our guide to free AI meeting notes.
FAQ
Is Ripple-65B free? You can download the weights for free under Apache 2.0. Hosted API use is paid.
Can I use it commercially? The Hugging Face repo lists the Apache 2.0 licence, which allows commercial use.
Does it really have 1M context? Ricursive says it supports up to 1M tokens. On OpenRouter, Prime Intellect and GMICloud offer 1,048,576 tokens; Together offers 131,072.
Is it better than Claude or GPT? We don’t know. The only benchmark results so far come from Ricursive.
Sources
- Ricursive, “Introducing Ripple-65B Preview”: https://www.ricursive.com/ripple (checked 2026-11-14)
- Hugging Face, ricursiveAI/Ripple-65B-Preview: https://huggingface.co/ricursiveAI/Ripple-65B-Preview (checked 2026-11-14)
- OpenRouter, Ripple-65B Preview: https://openrouter.ai/ricursiveai/ripple-65b-preview (provider table; checked 2026-11-14)
Last checked: November 14, 2026 (MYT).
1 thought on “Ripple-65B Preview: Free Open Model, $0.10/1M Tokens to Try”