Ricursive Intelligence, a startup better known for AI-driven semiconductor design, has released a general-purpose coding model called Ripple-65B Preview under the Apache 2.0 license. According to VentureBeat, the company says the model scores highest among U.S.-developed open-weights models on two third-party benchmarks: Terminal-Bench 2.1, which measures coding in a terminal environment, and SWE-bench Pro v2.0 Hard, a difficult software engineering suite introduced in September.

Availability and pricing

The weights can be downloaded from Hugging Face, and the model is served through API partners Together AI, GMI Cloud and OpenRouter, with Fireworks support expected. On OpenRouter, the GMI Cloud endpoint is priced at $0.20 per million input tokens and $0.60 per million output tokens. One discrepancy is notable: Ricursive advertises a context window of a million tokens, while OpenRouter currently lists a limit of 16,384 tokens, which could constrain real-world use until it is resolved.

Origins in chip design

Ripple is described as a general-purpose variant of the proprietary language model behind Ricursive's chip-design platform. Developers therefore get a related model, not the full chip-design system, and coding benchmark results say nothing about how well that separate platform works. CEO Anna Goldie and CTO Azalia Mirhoseini previously helped build Google's reinforcement-learning-based AlphaChip. The company 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.

The benchmark picture

Ripple has 65 billion parameters and was trained on 12 trillion tokens, with an emphasis on multistep, tool-using tasks. In Ricursive's launch comparison, it scored 80.5 on Terminal-Bench 2.1, narrowly ahead of 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. On SWE-bench Pro v2.0 Hard it posted 79.1 versus 76.5 for Beam and 72.4 for Meta. The company stresses that it reaches these results with roughly 15 times fewer total parameters than Inkling.

The lead does not extend across the board. Meta's model beats Ripple on DeepSWE (60.0 vs 44.3), on the FORTRESS adversarial safety test, and on AutomationBench, and Meta and Reflection both outscore it on the HMMT math test. Reflection's own published comparison also lists several Chinese-developed models above Ripple's Terminal-Bench figure, and its updated Beam numbers differ from those in Ricursive's table. Because these are separate vendor evaluations rather than a controlled head-to-head, they should be read cautiously.

Inconsistencies in the materials

Ricursive's Hugging Face model card lists SWE-bench Pro v2.0 Hard scores for Meta (52.3) and Inkling (2.0) that differ sharply from the launch graphic (72.4 and 39.6), and some DeepSWE rival scores also vary. The card acknowledges that its tables blend published and internal results, that agent benchmarks are hard to reproduce, and that changing the evaluation framework shifts scores. Since an agent's scaffolding determines tool access, long-context handling and failure recovery, benchmark numbers only partly predict performance in a team's actual workflow. Ripple is best seen as a candidate worth testing, not a proven replacement.

Efficiency claims

Ricursive says the entire effort, including experiments, data preparation, evaluations, synthetic data generation, later training stages and final runs, cost $32 million, far below the hundreds of millions or billions estimated for proprietary frontier models. It estimates spending about a tenth of what the previous U.S. open-weight coding leader, Reflection, spent, based on Reflection's reported compute commitments and training durations; this is an estimate, not a like-for-like disclosure. The startup also says its reinforcement learning reached the Terminal-Bench result after 1.4 million rollouts, about 70 times fewer than Reflection's approach, while Reflection reported more than 100 million rollouts in a four-week run. Rollout counts alone do not prove equivalent savings, since sequences differ in length, hardware demands and evaluation cost. Ricursive further credits gains in pretraining efficiency to tuning architecture, training settings and data, starting from an Olmo 3 baseline and testing changes at smaller scales first.

Why it matters

If the numbers hold up, Ripple shows that a small, well-funded team can approach larger open models in agentic coding at much lower cost and parameter count, adding pressure on Meta, Reflection, Thinking Machines and Nvidia in the U.S. open-weights race. For developers, a permissively licensed 65B model is cheaper to host and fine-tune than hundreds-of-billions-parameter alternatives. Independent reproduction and resolution of the context-window and scoring discrepancies will determine whether the claimed lead is durable.