Ricursive releases Ripple-65B, a US open-weight coding model trained for $32 million
The new 65-billion-parameter model sets agentic coding records at one-tenth the cost of prior state-of-the-art systems, according to the developer.
Key points
- Ripple-65B Preview achieves 80.52 on TerminalBench 2.1, surpassing larger US open-weight models.
- Ricursive reports total research and development costs of $32 million for the project.
- The model is available via Hugging Face, Together AI, GMI Cloud, OpenRouter, and Fireworks.
Ricursive Intelligence released Ripple-65B Preview, a language model that sets a new United States open-weight state-of-the-art on agentic coding benchmarks, according to a blog post published by the company.reported
What was released or claimed
The company stated that Ripple-65B Preview is a general-purpose variant of the proprietary model powering its AI chip design platform.reported Ricursive reported achieving a score of 80.52 on TerminalBench 2.1, outperforming significantly larger US open-weight models including Reflection Beam 501B, Meta Muse Spark Open 816B, Thinking Machines Inkling 975B, and NVIDIA Nemotron 3 Ultra 550B.reported The model supports up to one million tokens of context length and was pretrained on 12 trillion tokens.
Who it affects
Data science teams and developers seeking efficient local inference options may benefit from the model's dense architecture. Ricursive noted that the small parameter envelope enables efficient operation on local devices such as NVIDIA DGX Spark, which offer high compute FLOPs with low memory bandwidth relative to datacenter accelerators. The release targets users who require strong reasoning performance with low-latency token generation without relying on massive datacenter resources.
What is still unknown
The company did not disclose specific architectural details beyond describing the model as having a dense architecture. It also did not provide independent verification of the $32 million cost figure, noting it is an estimate of full R&D costs starting from scratch. The material does not specify whether the model's performance gains are reproducible across other agentic coding benchmarks or tasks outside of TerminalBench 2.1.reported
The takeaway
Ricursive claims that Ripple-65B Preview was developed for $32 million, which it describes as one-tenth the cost of the previous open-weight state-of-the-art on agentic coding.reported This figure includes pretraining ladder development, data mixing, evaluations, debugging, synthetic data generation, and final training runs. The company emphasized that it avoided buying posttraining data, instead using its own synthetic data pipelines and quality filtering methods.
What changes for users
Users can now access Ripple-65B Preview through Hugging Face, Together AI, GMI Cloud, and OpenRouter.reported Fireworks added support for the model later on the day of the announcement. Data science teams can deploy the model on local hardware like NVIDIA DGX Spark due to its optimized memory usage. The documented limitation is that this is a preview release, and independent audits of its agentic coding capabilities or cost efficiency have not been published.
Why it matters
- The $32 million R&D cost claim challenges prevailing industry estimates that frontier model training costs are rising toward billions of dollars.
- Surpassing larger models like Meta Muse Spark Open 816B on TerminalBench 2.1 suggests parameter count alone does not determine agentic coding performance.
- Availability on multiple inference platforms allows data science teams to benchmark the model against their existing pipelines without building custom infrastructure.
- Support for one million context tokens enables processing of larger codebases or documentation sets in single requests, improving workflow efficiency for large projects.
Topics: models, open-weight, agentic coding, Ricursive, AI training costs
Updated
Then / now / next
Previous US open-weight agentic coding leaders required significantly higher estimated costs and larger parameter counts.
Ripple-65B Preview offers higher benchmark scores at a reported fraction of the previous development costs.
Monitor independent evaluations of the model's performance on additional agentic coding tasks and real-world coding workflows.
Story timeline
- Then Previous US open-weight agentic coding leaders required significantly higher estimated costs and larger parameter counts.
- 04 Nov 2026 Ripple-65B Preview: U.S. Open-Weight SOTA on Agentic Coding
- Now Ripple-65B Preview offers higher benchmark scores at a reported fraction of the previous development costs.
- Next Monitor independent evaluations of the model's performance on additional agentic coding tasks and real-world coding workflows.
The strongest doubt
Counterpoint from the sources
The $32 million cost is an internal estimate by Ricursive and has not been independently audited or verified by third parties.
For readers
Developers
- Ripple-65B Preview is available on Hugging Face, Together AI, GMI Cloud, OpenRouter, and Fireworks for immediate testing.
- The model supports up to one million context tokens, enabling larger codebase analysis in single requests.
Researchers
- The model achieved 80.52 on TerminalBench 2.1, outperforming larger US open-weight models like Reflection Beam 501B.
- Ricursive utilized synthetic data pipelines and quality filtering instead of purchasing posttraining data.
Investors
- Ricursive claims total R&D costs of $32 million, significantly lower than industry estimates for frontier model training.
- The model is described as a preview variant of the proprietary system powering Ricursive's AI chip design platform.
Ask the article
Local keyword search of this article; no generated answers.Did this article give you the evidence you needed?
Your choice stays on this device. No public tally.Terms in this story
Agentic coding
The ability of AI models to autonomously perform complex coding tasks, including debugging and execution, rather than just generating code snippets.
Open-weight
Models whose parameters are publicly available for download and use, allowing for local deployment and modification.
TerminalBench 2.1
A benchmark suite used to evaluate the agentic coding capabilities of large language models.
Dense architecture
A model structure where all parameters are activated for every token, contrasting with sparse or mixture-of-experts designs.
Questions readers ask
How much did it cost to train Ripple-65B Preview?
Ricursive reported total research and development costs of $32 million, including all pretraining and posttraining phases.
Which benchmarks did Ripple-65B Preview top?
The model achieved a score of 80.52 on TerminalBench 2.1, setting a new US open-weight state-of-the-art record.
Can I run Ripple-65B Preview on local hardware?
Yes, the company states the model is optimized for local devices such as NVIDIA DGX Spark due to its dense architecture and small parameter envelope.
Reading receipt
- Ripple-65B Preview achieves 80.52 on TerminalBench 2.1, surpassing larger US open-weight models.
- Ricursive reports total research and development costs of $32 million for the project.
- The model is available via Hugging Face, Together AI, GMI Cloud, OpenRouter, and Fireworks.
