Key points
- Ripple-65B Preview scored 80.52 on TerminalBench 2.1, setting a new U.S. open-weight benchmark.
- The model outperformed significantly larger competitors including Reflection Beam 501B and Meta Muse Spark Open 816B.
- Total research and development costs were reported at 32 million dollars, including synthetic data generation.
Ricursive Intelligence released Ripple-65B Preview, claiming it sets a new U.S. open-weight state-of-the-art on agentic coding benchmarks.reported The company states the model achieved a score of 80.52 on TerminalBench 2.1.A new AI model claims to beat larger rivals at coding tasks while costing far less to build.Ricursive Intelligence asserts Ripple-65B achieves SOTA on TerminalBench 2.1 with dense architecture optimizations and synthetic data. This result surpasses larger models such as Reflection Beam 501B, which scored 80.1, and Meta Muse Spark Open 816B, which scored 76.4. The release is available on Hugging Face, Together AI, GMI Cloud, and OpenRouter, with Fireworks support expected later that day.
How we got here
Frontier AI training costs have risen sharply in recent years. OpenAI CEO Sam Altman stated in 2023 that GPT-4 training costs exceeded 100 million dollars. Dario Amodei later projected that frontier model costs could reach 1 billion dollars, potentially rising to 10-100 billion dollars by 2026 or 2027 for pretraining alone. These estimates highlight the financial barriers to entry for independent labs building competitive large language models from scratch.
What is new
Ricursive Intelligence reports developing and training Ripple-65B Preview for only 32 million dollars.reported This figure represents the full research and development cost, including pretraining ladder development, data mixing, evaluations, debugging, and synthetic data generation. The company notes this is one-tenth the cost of the previous open-weight state-of-the-art on agentic coding, according to their estimates.reported The model uses a dense architecture optimized for low-latency token generation and efficient local device inference.
Reading between the lines
The lab avoided buying post-training data, instead creating its own synthetic data pipelines and quality filtering metrics. They iterated on supervised fine-tuning and reinforcement learning recipes with a focus on compute efficiency rather than brute-force scaling. The model supports up to 1 million context length and was pretrained on 12 trillion tokens. Its small parameter envelope allows efficient operation on devices like NVIDIA DGX Spark, which have high compute but lower memory bandwidth.
The open questions
The claim of state-of-the-art performance rests on TerminalBench 2.1 scores, which may not reflect broader general-purpose capabilities. The 32 million dollar cost figure is an internal estimate including all R&D phases, not just compute, making direct comparisons to other labs' reported training costs difficult.reported The sources acknowledge these are preview results from a proprietary model variant released as open-weight. Independent verification of the cost breakdown and benchmark reproducibility remains pending.
Why it matters
- Evaluators can test whether agentic coding performance scales linearly with parameter count or if architectural efficiency drives results.
- The low cost claim challenges assumptions about the capital required to compete with major tech firms in open-weight models.
- Synthetic data pipeline details offer a template for labs seeking to reduce reliance on purchased post-training datasets.
- Local inference on hardware like DGX Spark expands the testing environment for model latency and memory efficiency metrics.
Topics: models, open-weight, agentic coding, benchmarks, Ricursive Intelligence
Updated
Then / now / next
Frontier AI training costs were projected to reach billions of dollars by 2026.
Ricursive Intelligence claims a new SOTA model was developed for 32 million dollars.
Fireworks support for the model goes live later that day.
Story timeline
- Then Frontier AI training costs were projected to reach billions of dollars by 2026.
- 04 Nov 2026 Ripple-65B Preview: U.S. Open-Weight SOTA on Agentic Coding
- Now Ricursive Intelligence claims a new SOTA model was developed for 32 million dollars.
- Next Fireworks support for the model goes live later that day.
The strongest doubt
Counterpoint from the sources
The 32 million dollar figure includes all R&D, not just compute, making direct cost comparisons with other labs' reported training expenses difficult.
Numbers in context
Circle area is proportional to the magnitude; labels retain signs. Figures may describe different measures. Read their contexts above.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
Automated software development tasks performed by AI agents.
TerminalBench 2.1
A benchmark suite for evaluating AI coding agent performance.
Synthetic data
Artificially generated data used for model training.
Open-weight
Model weights released for public use and modification.
Questions readers ask
How much did Ripple-65B cost to train?
Ricursive Intelligence reports total R&D costs of 32 million dollars, including all development phases.
Which benchmarks does the model lead?
It claims U.S. open-weight state-of-the-art on TerminalBench 2.1 with a score of 80.52.
Where can I access the model?
It is available on Hugging Face, Together AI, GMI Cloud, and OpenRouter, with Fireworks going live later that day.
Connected stories
Reading receipt
- Ripple-65B Preview scored 80.52 on TerminalBench 2.1, setting a new U.S. open-weight benchmark.
- The model outperformed significantly larger competitors including Reflection Beam 501B and Meta Muse Spark Open 816B.
- Total research and development costs were reported at 32 million dollars, including synthetic data generation.
Still open: Has the 32 million dollar cost estimate been independently audited?
Still open: Do other benchmarks show similar performance advantages over larger models?
