Ricursive Releases Ripple-65B, an Open Model for Coding Agents
A five-person lab says its 65B model beats open rivals with 500B or more parameters on TerminalBench. The scores are its own, so test before you switch.
Ricursive Intelligence released Ripple-65B Preview on November 4, an open-weight model that the lab says beats far bigger U.S. open-weight models at agentic coding.
Do you run coding agents for your sites or your product? Then this is one more model you can run on your own hardware, without sending your code to anyone else’s API.
What Ricursive released
Ripple-65B is a dense model with 65 billion parameters, pretrained on 12 trillion tokens, with up to 1M tokens of context.
Dense matters here.
The rivals it lines up against are mixture-of-experts models with 501B to 975B total parameters. Only 23B to 55B of theirs are active at once, while all 65B of Ripple’s are.
The Ripple-65B model card on Hugging Face lists three versions, and the main repo carries an Apache 2.0 licence:
- the instruction-tuned model in BF16
- the same model in NVFP4, which the card offers for local use
- a BF16 base model, with no extra training on top
You don’t have to download it, either: Ricursive says it’s also on Together AI, GMI Cloud and OpenRouter, and that Fireworks would follow later that day.
The scores, and who ran them
The headline claim is a TerminalBench 2.1 score of about 80.5. Here’s how Ricursive says the U.S. open-weight models compare at agentic coding on that test:
| Model | Total parameters | TerminalBench 2.1 |
|---|---|---|
| Ripple-65B Preview | 65B | 80.5 |
| 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 |
Ripple’s 80.5 is Ricursive’s own result, from its own test runs.
Where a lab didn’t publish a score for one of the tests, the card fills the gap with Ricursive’s in-house runs and calls the whole table “a best-effort attempt at a fair comparison.”
But it doesn’t win everywhere.
On SWE-bench Verified it scores 76.5 against Beam’s 81.6. On MCP Atlas, Muse Spark Open leads by about 89 to 72.
Then there’s the cost. Ricursive puts its full R&D bill at $32M, from scratch, which it estimates is about a tenth of what the previous open-weight leader in agentic coding cost to make.
Its team grew from one person to five during the build, the card says, and an internal variant of the same model powers Ricursive’s own AI chip design platform.
Why it matters
So who should care? I’d put three groups first:
- SaaS founders whose product gets written, tested and shipped with the help of coding agents every week
- agencies running client sites with AI-written code
- anyone who’d rather keep their code, and their customers’ data, on their own hardware
My read: open-weight models take away two worries at once. Your code stays where you choose, and nobody can change your model’s price or rules overnight.
I think the gap between the models you rent by the token and the models you run on your own hardware is closing fast, and a 65B model even claiming to top 500B rivals is a sign of it.
Ripple lands after Kimi K3.1 Pro’s open weights and DeepSeek’s open-source Super-agent.
Thinking of running it at home? Check your memory first, because the BF16 files alone come to about 130 GB and the card names one example machine for local use, NVIDIA’s DGX Spark.
Still, I wouldn’t move a live workflow onto a preview when the card itself says post-training was relatively light and focused on agentic coding.
So test it the cheap way first:
- Run it on OpenRouter with three or four real tasks from your own repo, the ones your current model already handles well.
- Compare the diffs and count every task that needs a second try.
- Then try the NVFP4 checkpoint at home.
That said, a five-person lab matching 500B models is news.
If outside tests agree, you get a coding model you can run on your own hardware. If not, you’ve lost an afternoon.
Common Questions
Can I use Ripple-65B Preview in a commercial product?
Yes. The Ripple-65B model card on Hugging Face shows an Apache 2.0 licence for the main repo, which allows commercial use. Check the licence file of the version you download.
Can I run Ripple-65B on my own hardware?
Only on serious hardware. Ricursive says Ripple-65B fits on consumer hardware such as an NVIDIA DGX Spark, and it offers an NVFP4 version for local use, while the BF16 weights alone take about 130 GB.
Is Ripple-65B better than Claude or GPT for coding?
Ricursive doesn’t say. It compares Ripple-65B only with U.S. open-weight models at agentic coding. Its TerminalBench 2.1 score of about 80.5 is Ricursive’s own result, from its own test runs.
