Ripple-65B reaches 80.52 on TerminalBench 2.1

What changed

Ricursive Intelligence released Ripple-65B Preview on November 6, 2026. The model achieved a score of 80.52 on TerminalBench 2.1, claiming the top rank among US open-weight models for agentic coding. Ripple-65B was pretrained on 12 trillion tokens and supports up to 1 million tokens of context length. The model is available on Hugging Face, Together AI, GMI Cloud, OpenRouter, and Fireworks under the Apache 2.0 license.

The architecture uses a dense Transformer with 65 billion parameters instead of a mixture-of-experts approach. According to Ricursive, the full development and training effort cost $32 million, covering all research, development, pretraining, and posttraining expenses from the beginning.

Ripple surpasses several larger US open-weight models on TerminalBench 2.1: Reflection Beam 501B (80.1), Meta Muse Spark Open 816B (76.4), Thinking Machines Inkling 975B (63.8), and NVIDIA Nemotron 3 Ultra 550B (56.4). However, on other benchmarks, the picture differs. Inkling 2 Small scores higher on SWE-Bench Pro v2.0 Hard (86.3% vs. 79.1%), DeepSWE v1.1 (72.1% vs. 44.3%), and MCP Atlas (85.4-85.0% vs. 71.67%).

Why it matters

Open source models optimized for coding allow teams to run agentic systems on their own hardware without relying on external APIs or proprietary services. Ripple-65B’s performance on TerminalBench 2.1 puts it ahead of larger US-based open models, which matters for cost-conscious teams building coding agents. For deployment, this means access to a capable model that runs efficiently on hardware like NVIDIA DGX Spark, which offer high compute throughput with lower memory bandwidth than typical datacenter setups.

The cost story signals a potential shift in model development economics. If Ricursive’s $32 million development budget truly represents the frontier for open agentic coding at this scale, it suggests that careful architecture choices and efficient training recipes can compete with much higher spending. This could influence how other labs approach model development and whether open source remains viable as compute budgets grow.

The dense architecture choice carries deployment implications. Mixture-of-experts models often require more memory bandwidth and infrastructure complexity. Ripple’s thin-and-deep design, with a narrow model dimension of 5120 and 80 layers, was tuned to train and serve efficiently on NVIDIA hardware. This may appeal to teams with standard cloud or on-premises GPU clusters.

Ricursive’s posttraining approach also differs from industry norms. Rather than purchasing posttraining data, the company developed synthetic data pipelines with quality filtering and coverage metrics. The RL recipe reached a TerminalBench 2.1 score of 80.5 using only 1.4 million rollouts, roughly 70 times fewer than Reflection Beam according to the company’s estimate.

What to test

Before adopting Ripple-65B for production, practitioners should verify several claims:

• Benchmark isolation: The 80.52 TerminalBench 2.1 score compares to other models on the same benchmark, but test Ripple on your own agentic coding tasks to confirm real-world performance aligns with reported numbers. Performance varies significantly across benchmarks.
• Cost claims: Ricursive states it reached an 80.5 score with only 1.4 million RL rollouts, roughly 70 times fewer than Reflection Beam. Independently verify that latency, throughput, and total cost per inference match the company’s efficiency claims in your deployment scenario.
• Memory and compute requirements: The model is optimized for NVIDIA hardware. Test it on your specific GPU cluster or cloud setup to measure actual memory footprint, token generation latency, and throughput under load.
• Context length utility: The model supports 1M context tokens, but confirm that your use case actually benefits from such length and that quality does not degrade at the extremes.
• License compliance: Apache 2.0 is permissive, but verify that your intended use aligns with the license terms and any service provider agreements if you use a hosted version.
• Benchmark comparison caveats: Ripple ranks significantly below Inkling 2 Small on DeepSWE v1.1 (44.3% vs. 72.1%) and scores lower on SWE-Bench Pro v2.0 Hard (79.1% vs. 86.3%). Understand which benchmarks matter most for your problem.

Common questions

What is the strongest open-source coding model?

Performance depends on which benchmark matters for your use case. Ripple-65B Preview achieves the best result among US open-weight models on TerminalBench 2.1 with a score of 80.52. However, Inkling 2 Small scores higher on DeepSWE v1.1, SWE-Bench Pro v2.0 Hard, and MCP Atlas. The “strongest” model depends on your specific task.

Are there any truly open source AI models?

Yes. Ripple-65B Preview is released under the Apache 2.0 license, which permits commercial and private use with minimal restrictions. The weights and model are available on Hugging Face and other open repositories, making it a genuinely open source option for coding.

What are the main trade-offs between Ripple-65B and larger models?

Ripple-65B uses a dense architecture with 65 billion parameters, whereas competitors like Reflection Beam (501B) and Meta Muse Spark Open (816B) are much larger. Ripple’s smaller size means lower memory requirements and faster token generation on standard hardware, but it scores lower on some tasks like DeepSWE v1.1. Choose based on your latency, cost, and accuracy requirements.

The conclusion

Ripple-65B Preview represents a pragmatic approach to open source coding models: strong benchmark performance on TerminalBench 2.1 achieved through careful architecture design and efficient training rather than raw scale or brute-force compute spending. The Apache 2.0 license and availability across multiple inference platforms give teams flexibility to deploy locally or via hosted services.

However, practitioners should not assume Ripple leads across all coding tasks. Its TerminalBench 2.1 lead does not carry over to DeepSWE or SWE-Bench Pro, where Inkling 2 Small performs substantially better. Teams should test Ripple on their actual workflows before committing.

The cost claim of $32 million for frontier performance is noteworthy if accurate, but remains difficult to verify independently. Monitor how other labs respond and whether they can replicate similar efficiency gains. If they do, open source coding models may remain competitive even as frontier labs push toward billion-dollar training budgets.

Watch for updates on how Ripple scales beyond 65 billion parameters and whether the efficiency gains from the model ladder transfer to larger variants. Also track real-world adoption and whether the claimed latency and throughput advantages hold under production load.

Related reading

• Letterpress-1: Coding agent for growth engineering
• NVIDIA Auto Ontology: Open semantic layer for AI agents
• dsh-superagent: Router-worker coding agent for parallel GitHub tasks