Ripple-65B: an open-weight coding model from the creators of AlphaChip

On November 4, Ricursive Intelligence released Ripple-65B Preview, the company’s first public model and a Ripple-65B open-weight model on Hugging Face under the Apache 2.0 license. It has 65 billion parameters, it is built for agentic coding and, according to Ricursive, it scores 80.5 on Terminal-Bench 2.1.

What is the Ripple-65B open-weight model?

It is a dense language model, so no MoE (mixture-of-experts): all 65 billion parameters work on every token. Ricursive trained it from scratch on 12 trillion tokens (mostly English text and code), and the context window goes up to 1 million tokens.

Interestingly, Ricursive is not a typical chatbot lab. The company builds an AI platform for chip design, and it was founded by Anna Goldie and Azalia Mirhoseini, the creators of AlphaChip. The announcement calls Ripple a “general-purpose variant”, a general version of the closed model that powers their chip design platform.

(Don’t confuse it with Recursive Flint: that’s a different company and a different model, even though the names sound almost the same ;))

What can Ripple-65B do?

According to the model card, post-training focused on agentic coding. The most interesting results reported by Ricursive:

1. 80.5 on Terminal-Bench 2.1, a test in which the model performs tasks in a terminal. The announcement calls this “state-of-the-art” among American open-weight models.
2. 76.5% on SWE-bench Verified and 61.42% on SWE-bench Pro v1.0, which measure fixing real issues from repositories.
3. 85.2% on GPQA Diamond (PhD-level science questions) and 95.0% on AIME 2026 (math).
4. 71.67% on MCP Atlas, a test of tool use over MCP.

Let’s take these numbers with a grain of salt. The model card says many results come from Ricursive’s own harnesses, and it describes the whole Preview release as one where post-training was “relatively light”. There are no independent measurements yet.

Why is everyone talking about it?

First, the size. 65 billion parameters is not much for a model aiming at the top of agentic coding, and according to the announcement the model can “run efficiently on local devices such as NVIDIA DGX Spark”. For running it locally there is a ready NVFP4 (quantized) checkpoint and a Colab notebook.

Second, the cost. According to the announcement, the entire development and training of Ripple-65B Preview cost $32 million, and the Terminal-Bench result took only 1.4 million rollouts of RL (reinforcement learning).

Third, the architecture. The model is “thin and deep” (80 layers, d_model 5120), and instead of the typical Llama-style design it uses techniques similar to Gemma 3N, including gated attention and value residuals. According to the model card, this change improved training stability.

On Hacker News, the launch post made the front page and collected over 140 points within a few hours.

How to try Ripple-65B?

There are a few ways:

1. Locally: the BF16 weights, the NVFP4 version and the base model are on Hugging Face. If you are just getting started with models on your own hardware, the guide on how to run AI locally will help.
2. Via API: the model is available on Together AI, GMI Cloud and OpenRouter, and Fireworks plans to add it soon. On OpenRouter it costs $0.20 per 1M input tokens and $0.60 per 1M output tokens (check the provider’s context limit: on OpenRouter it is currently 16K tokens, not 1M).

According to the model card, on two B200 GPUs the model generates about 250 tokens per second (at batch size 8), still without quantization or custom kernels.

What to watch out for?

First, the Preview status. Ricursive hasn’t said when the full version will arrive. The announcement only mentions capabilities that are to come “as we scale toward the frontier”.

Second, the language. Pretraining was mostly in English and on code, so the model may do worse in other languages than it does in the terminal.

This is another strong open-weight model in recent weeks, after Reflection Beam and Kimi K3.1 Pro. This time, though, with a version you can set up on your desk.

FAQ

Is Ripple-65B free?

Yes, the weights are available under the Apache 2.0 license, so we can download the model and use it commercially too. We only pay for hardware and electricity, or for an API provider.

Who created Ripple-65B?

Ricursive Intelligence, a Palo Alto company building an AI platform for chip design, founded by Anna Goldie and Azalia Mirhoseini.

Is Ripple-65B a MoE model?

No. It is a dense model with 65 billion parameters, all of which are active for every token.