Ricursive Intelligence, the chip-design startup founded by the researchers behind Google's AlphaChip, has released Ripple-65B Preview, a 65-billion-parameter open-weight AI model it says is now the best American open model at agentic coding — built from scratch for just $32 million.
The model, released on November 4 under the Apache 2.0 licence, is available now on Hugging Face and through API providers including Together AI, GMI Cloud and OpenRouter, with Fireworks support to follow.
A small model beating giants
Ricursive puts Ripple-65B at 80.52 on TerminalBench 2.1, a benchmark that scores AI agents on executing real coding tasks inside a Linux terminal. That edges out Reflection's Beam at 80.1, Meta's Muse Spark Open at 76.4, Thinking Machines' Inkling at 63.8 and Nvidia's Nemotron 3 Ultra at 56.4 — rivals carrying between 501 billion and 975 billion total parameters, up to 15 times Ripple's size.
On SWE-bench Pro v2.0 Hard, which covers difficult software engineering tasks, Ripple scores 79.1, again ahead of the same group. The model is dense, meaning all 65 billion parameters are active on every token, and it was pretrained from scratch on 12 trillion tokens of mostly English text and code, with a context window of up to one million tokens.
Ripple is a general-purpose variant of the proprietary model that powers Ricursive's AI chip-design platform, built on the Allen Institute for AI's open Olmo 3 recipe.
$32 million and 1.4 million practice runs
The headline number may be less striking than the price tag. Ricursive says the $32 million figure is its full research and development bill — the small test models, data work, evaluations, debugging, synthetic data and the final training runs — not just the cost of the last run. The company estimates that is about a tenth of what the previous open-weight leader spent.
The efficiency extends to training. Ripple reached its TerminalBench score with about 1.4 million reinforcement learning rollouts, roughly 70 times fewer than the 100 million-plus reported by Reflection's Beam. Ricursive says it bought no post-training data, generating its own synthetic data instead.
On OpenRouter the model costs $0.20 per million input tokens and $0.60 per million output tokens. Its model card claims 250 tokens per second on two Nvidia B200 chips.
The fine print
Every benchmark and cost figure here is the company's own, and none has been independently replicated yet. Ricursive itself warns that agentic benchmark scores are hard to compare: some rival numbers come from each lab's own model card and some from its own test runs, with different evaluation harnesses. When it reran TerminalBench in the official Terminus 2 harness, Beam's score fell to 70.8 while Ripple's held at 80.5 — but independent checks are still to come.
Its own results also show Ripple trailing the larger models on several other tests, including knowledge, maths and science coding, and there are discrepancies within Ricursive's materials: its model card lists different SWE-bench Pro v2.0 Hard scores for rivals than its launch graphic does. The advertised million-token context also exceeds the 16,384-token limit currently listed by OpenRouter.
Still, the release underlines how quickly the open-model field is compressing. A chip-design startup that raised $335 million at a $4 billion valuation in four months, according to TechCrunch, has shipped a model that — on its own numbers — out-codes rivals many times its size for a tenth of the cost. Whether independent testing confirms the claim will decide if Ripple is a genuine efficiency breakthrough or just a well-aimed benchmark run.



