A five-person team at a chip-design startup says it built one of the best open coding models in the world for $32 million, and the numbers it published on November 4 are hard to ignore. Ricursive Intelligence released Ripple-65B Preview under the Apache 2.0 licence, free to download and use commercially, and claims it beats American open-weight rivals up to 15 times its size at agentic coding.
The claim comes with a long list of footnotes, and some of them matter more than the headline. But even the sceptical reading leaves a striking result: a small lab, founded by the researchers behind Google’s AlphaChip, trained a 65-billion-parameter model from scratch and put it within touching distance of models that cost far more to build.
A 65-billion-parameter model trained on 12 trillion tokens
Ripple-65B Preview is a dense Transformer, which means all 65 billion of its parameters are active on every token, unlike the mixture-of-experts designs that dominate the frontier. Ricursive trained it from zero on 12 trillion tokens, mostly English prose and source code, and says it handles up to 1 million tokens of context.
The lab is not an AI company by trade. Ricursive builds AI for chip design, and a private variant of the model already powers its chip-design platform. The public model is a general-purpose release, available now on Hugging Face and through Together AI, GMI Cloud and OpenRouter, with Fireworks support expected soon. The Apache 2.0 licence means anyone can download the weights, run them locally, or build a product on top of them without paying Ricursive a cent.
Terminal-Bench 2.1: ahead of Reflection Beam by 0.4 points
The headline result is a score of 80.52 on Terminal-Bench 2.1, a test that asks an AI agent to complete real coding jobs inside a terminal. Ricursive calls it the best score among US open-weight models, and the comparison table it published makes the case:
| Model | Parameters | Terminal-Bench 2.1 | SWE-bench Pro v2.0 Hard |
|---|---|---|---|
| Ripple-65B Preview | 65B | 80.5 | 79.1 |
| Reflection Beam | 501B | 80.1 | 76.5 |
| Meta Muse Spark Open | 816B | 76.4 | 72.4 |
| Thinking Machines Inkling | 975B | 63.8 | 39.6 |
| Nvidia Nemotron 3 Ultra | 550B | 56.4 | 30.7 |
The lead on Terminal-Bench is a hair, 0.4 points over Reflection Beam, but the SWE-bench Pro Hard margin is clearer, and the size gap is the real story. Ripple has roughly 15 times fewer parameters than Thinking Machines’ Inkling and still outscores it by nearly 17 points on the terminal test.
The $32 million build, and the 1.4 million practice runs
Ricursive says $32 million covers the entire research and development effort, not just the final training run. That includes the small test models, data work, evaluations, debugging, synthetic data and the last big training stages. The lab says it bought no post-training data and generated its own synthetic data instead.
The team grew from one person to five during the project, and it reached its Terminal-Bench score with about 1.4 million reinforcement-learning rollouts, which it says is roughly 70 times fewer than Reflection Beam used. If the figure holds up, it suggests the expensive part of building a good coding model is no longer the raw compute, but knowing which experiments to run.
Context helps here. Ricursive raised $335 million at a $4 billion valuation in four months earlier this year, according to TechCrunch, so the lab is small in headcount but not short of money. The $32 million is the cost of this model, not the cost of the company behind it.
Where Ripple loses, and why the fine print matters
Ricursive’s own model card shows Ripple is not a clean sweep. On SWE-bench Verified, the standard bug-fixing test, it scores 76.5 against 81.6 for Reflection Beam. On MCP Atlas, a tool-calling test, it posts 71.67 against 89.2 for Meta’s Muse Spark Open. On SciCode, a science-coding test, it gets 39.4 against that same model’s 58.5.
Every number in the launch comes from Ricursive, and no independent lab has reproduced the results yet. The model card is candid about the difficulty: some rival scores come from published model cards and some from Ricursive’s own test runs, using different harnesses, and the card marks which is which. One discrepancy is worth noting. Under the official Terminus 2 harness, Reflection Beam’s Terminal-Bench score drops from 80.1 to 70.8, while Ripple’s stays at 80.5. That could mean Ripple is genuinely more robust, or it could mean the harness change hits Beam’s setup harder. Nobody outside Ricursive can say yet.
There is also a version trap waiting for anyone comparing Ripple with closed models. Ripple’s 80.5 is on Terminal-Bench 2.1. Anthropic’s Claude Opus 5.5 scores 66.4 on Terminal-Bench 4.0, a newer and much harder version of the test, where scores for the same models dropped by 20 to 70 points. The two numbers are not comparable, and anyone putting them in the same column is making a mistake.
What it costs to use, and what to watch next
For developers, the practical numbers are the hosted prices. OpenRouter lists Ripple at around $0.20 per million input tokens and $0.60 per million output tokens, a fraction of the $4 input and $20 output that Claude Opus 5.5 charges. The weights themselves are free, so anyone with a capable GPU can run the model without an API bill at all.
One claim deserves a check before you plan around it. Ricursive advertises a 1-million-token context window, but OpenRouter currently lists a 16,384-token limit for the model, so the full window may not be reachable through every provider yet.
The honest verdict is that Ripple-65B Preview is a genuinely interesting release with a genuinely unproven headline. The $32 million cost, the five-person team and the small-model-beats-giants table are all Ricursive’s own claims, and the first independent Terminal-Bench runs will decide whether the 80.52 survives outside the lab’s own harness. But the direction of travel is clear. A month ago, Moonshot gave away Kimi K3.1 Pro and Mistral open-sourced a 1-trillion-parameter model, and now a five-person startup says it built a top-tier coding agent for the price of a mid-size office building. The moat around frontier coding models is getting shallower, and the next test is whether anyone else can repeat the trick.
Sources: Ricursive Intelligence launch post and Hugging Face model card, November 4, 2026; VentureBeat; OpenRouter listing.
