On 4 November a small lab called Ricursive Intelligence released Ripple-65B Preview: a dense 65-billion-parameter model trained from scratch on 12 trillion tokens, post-trained for agentic coding, and published under Apache-2.0. Its headline number, 80.5 on Terminal-Bench 2.1, is ahead of open mixture-of-experts models with ten times more parameters. Running it yourself takes a GPU with 40 GB or more, so we host it: you can try it in your browser.
Ripple-65B PreviewRicursive Intelligence · Apache-2.0A new 65B dense model trained from scratch for agentic coding: 80.5 on Terminal-Bench 2.1, ahead of open models ten times its size.Try it online →What Ripple-65B is
- 65B dense (not mixture-of-experts): every token uses all the weights, which makes it simple to serve and quantize.
- Thin and deep: 80 layers with a narrow 5,120 width, mostly sliding-window attention with a few global layers, in a Gemma-3n-like design.
- Long context: trained up to 1M tokens (on LiveWeights each chat uses a 16K window).
- Reasoning model: on harder questions it thinks before answering. On LiveWeights you can open its reasoning under each answer.
- Built by a team of five, starting from the open Olmo 3 recipe.
Benchmarks, as reported by Ricursive
| Benchmark | Ripple-65B Preview |
|---|---|
| Terminal-Bench 2.1 | 80.5 |
| SWE-bench Verified | 76.5 |
| SWE-bench Pro v2.0 Hard | 79.1 |
| GPQA Diamond | 85.2 |
| AIME 2025 | 98.3 |
| MMLU-Pro | 84.2 |
These are the lab's own numbers at full BF16 precision. It is a "Preview", with light post-training focused on coding: expect it to be strongest on code, terminal tasks and reasoning, and weaker on general knowledge than much bigger models.
How to try it online
- Open the Ripple-65B page and sign in, then pick a small credit pack (no subscription).
- Ask a coding question, paste a bug or ask for a script.
- The first message after a pause needs about 40 seconds to load the model on our 96 GB GPU; after that it writes about 34 tokens per second.
We run a 4.6-bit GGUF build with llama.cpp, which added Ripple's architecture on 5 November. Quality is close to the original, but not identical: for benchmarking, use the official BF16 weights.
Run it locally
You need about 40 GB of GPU memory (or unified memory on a Mac or a DGX Spark) for a 4-bit build. Options: the official NVFP4 checkpoint for NVIDIA Blackwell, GGUF builds for llama.cpp (use a release from 6 November 2026 or later, and a GGUF converted after the merge), or MLX builds for Apple Silicon. For other new open chat models, see this week's roundup.