Ricursive's Ripple-65B Preview tops US open models on agentic coding — at a claimed $32M build cost
Ricursive Intelligence's 65-billion-parameter Ripple-65B Preview claims the top US open-weight scores on two agentic coding benchmarks, beating models up to fifteen times its size — with a build cost the startup puts at $32 million. The figures are the company's own, and the fine print matters.
- The model: a dense 65B-parameter model trained from scratch on 12 trillion tokens, with a 1M-token context window; Apache 2.0 weights on Hugging Face, plus hosted access via Together AI, GMI Cloud and OpenRouter.
- The scores: 80.52 on Terminal-Bench 2.1 and 79.1 on SWE-bench Pro v2.0 Hard, per Ricursive — ahead of Reflection Beam, Meta Muse Spark Open, Thinking Machines Inkling and Nvidia Nemotron 3 Ultra on both. It trails on DeepSWE and other tests in the same table.
- The cost: $32 million end to end, by Ricursive’s own accounting — about a tenth of what it estimates the previous US open-weight coding leader spent. No independent audit.
- The catch: all figures are vendor-reported and not independently reproduced; Ricursive’s own materials disagree on some rival scores, and OpenRouter’s endpoint caps context at 16,384 tokens despite the advertised 1M window.
- The backstory: Ricursive is a chip-design startup founded by Anna Goldie and Azalia Mirhoseini, creators of Google’s AlphaChip. Ripple is a general-purpose variant of the model behind its chip-design platform.
A chip-design startup that most people outside semiconductors had never heard of just posted the best American open-weight scores in agentic coding — with a model one-fifteenth the size of the rival it beat. Ricursive Intelligence released Ripple-65B Preview on November 4, a 65-billion-parameter model it says cost $32 million to build end to end. The claim is striking; the evidence, as the company itself concedes, is its own.
What Ripple-65B Preview actually is#
Ricursive Intelligence is a chip-design startup founded by Anna Goldie and Azalia Mirhoseini, the two researchers behind Google’s AlphaChip system for reinforcement-learning chip design. Ripple is the company’s first public language model: a general-purpose variant of the proprietary model that powers its chip-design platform.
The model is dense: 65 billion parameters, all of them active on every token, trained from scratch on 12 trillion tokens with a context window of up to 1 million tokens. It ships under the Apache 2.0 license, which allows commercial use, with weights on Hugging Face and hosted access through Together AI, GMI Cloud and OpenRouter, with Fireworks to follow. Three checkpoints are published: an instruction-tuned BF16 build, a compressed NVFP4 build aimed at local hardware, and a pretrain-only base model. The “Preview” label is the company’s own warning: post-training was light, and the instruct model is a first cut.
The timing is part of the story. Ricursive raised a $300 million Series A in January at a $4 billion valuation, led by Lightspeed with Nvidia’s NVentures, DST Global and Sequoia participating. Its main business is still chip design; Ripple is a way to let outsiders judge the technology behind that platform.
The scores — and the fine print#
Ricursive’s headline claim is that Ripple leads every US open-weight model on two agentic coding benchmarks: 80.52 on Terminal-Bench 2.1, a test of agents doing real work in a Linux terminal, and 79.1 on SWE-bench Pro v2.0 Hard, a suite of difficult software-engineering tasks. On Terminal-Bench it edges Reflection’s Beam (80.1) and sits clearly ahead of Meta’s Muse Spark Open (76.4), Thinking Machines’ Inkling (63.8) and Nvidia’s Nemotron 3 Ultra (56.4). On SWE-bench Pro Hard the gap is wider: Beam 76.5, Muse Spark 72.4, Inkling 39.6, Nemotron 30.7.
| Benchmark | Ripple-65B Preview (65B) | Reflection Beam (501B) | Meta Muse Spark Open (816B) | Thinking Machines Inkling (975B) | Nvidia Nemotron 3 Ultra (550B) |
|---|---|---|---|---|---|
| Terminal-Bench 2.1 | 80.5 | 80.1 | 76.4 | 63.8 | 56.4 |
| SWE-bench Pro v2.0 Hard | 79.1 | 76.5 | 72.4 | 39.6 | 30.7 |
| DeepSWE 1.1 | 44.3 | 43.3 | 60.0 | 5.0 | 0.6 |
The same launch materials show the lead is narrow and selective. On DeepSWE, Ripple scores 44.3 against Muse Spark’s 60.0. Meta also leads on the displayed automation and adversarial-safety tests, and Meta and Reflection both score higher on the math test. Reflection’s own published comparison lists several Chinese-developed models above Ripple’s Terminal-Bench result. And Ricursive’s own numbers disagree with each other: its model card lists SWE-bench Pro Hard scores of 52.3 for Meta and 2.0 for Inkling, against 72.4 and 39.6 in the launch graphic. None of these figures has been independently reproduced, and agent benchmarks are notoriously sensitive to the harness around the model — the tools it can call, how it manages long conversations, how it handles failure.
How Ricursive says it spent only $32 million#
The other headline number is the cost. Ricursive says the full research and development bill — early experiments, data work, evaluations, debugging, synthetic data and the final training runs — came to $32 million, about a tenth of what it estimates the previous US open-weight coding leader spent. It credits two things: a “model laddering” approach that ran systematic small-scale experiments, starting from the Allen Institute for AI’s open Olmo 3 recipe, before committing to the big run; and a reinforcement-learning process that reached its Terminal-Bench result after 1.4 million rollouts, roughly 70 times fewer than Reflection’s reported 100 million-plus.
The efficiency claims deserve the same scrutiny as the scores. The $32 million is Ricursive’s own accounting, and the one-tenth comparison is built from Reflection’s reported compute commitments and training durations — not matched disclosures from both companies. Rollout counts don’t map directly onto cost either: action sequences differ in length, hardware and evaluation expense. What is verifiable is the design choice behind the efficiency pitch: a dense architecture, which Ricursive says simplified its training infrastructure versus routing work through slices of a much larger mixture-of-experts model, and a reported roughly 40% model FLOPs utilization during training without custom kernels or quantization.
Small enough to run at home#
The practical appeal is deployment. At 65 billion parameters, Ripple is small enough to run on Nvidia’s DGX Spark desktop machine — the model card cites 250 tokens a second on two B200 chips — and the NVFP4 checkpoint is aimed at local hardware. Ricursive’s announcement also names MacBooks, though it establishes no specific laptop configuration, memory or speed for that claim. Hosted, OpenRouter’s GMICloud endpoint charges $0.20 per million input tokens and $0.60 per million output tokens, with cached input at $0.05. One catch: that endpoint currently supports 16,384 tokens of context, far short of the advertised million-token window — so buyers shouldn’t assume the price buys the full context.
What to watch#
Ripple’s real test starts now. Independent evaluations will show whether the scores survive outside Ricursive’s harness, and whether a model that is cheap to train is also cheap per finished task — token prices say nothing about how much text an agent generates, how often it retries, or how much human intervention it needs. Watch too for the full release: the Preview label means the weights are a first cut, and Ricursive’s chip-design platform — the business the model is meant to advertise — is judged by silicon, not benchmarks. If the numbers hold up, though, the message is hard to miss: the open-model race is no longer only about who can train the biggest model, but who can train the smallest one that still wins.
Sources: VentureBeat’s launch coverage, venturebeat.com; and MadRobot’s report, madrobot.blog. All benchmark figures and the $32 million cost figure are as reported by Ricursive Intelligence and have not been independently reproduced. The cover and inline illustrations are AI-generated.