Ripple-65B: an Apache 2.0 coding model you can run locally

A 65-billion-parameter open coding model just posted agent scores that used to need a model several times its size. Ricursive Intelligence released Ripple-65B Preview under Apache 2.0, and its self-reported Terminal-Bench 2.1 score of 80.5 beats several open models with hundreds of billions of parameters. The size is the story for a small team. It's built to run on hardware you can own.

What actually happened

Ricursive published Ripple-65B Preview on November 6 (UTC), with weights on Hugging Face. Per the model card and launch post:

• Dense, 65B. Every parameter is active on every token. No mixture-of-experts.
• Apache 2.0. Commercial use, modification and redistribution are allowed, with no separate agreement.
• Variants. An instruction-tuned model in BF16 and NVFP4, plus a pretrain-only base model.
• Context. Up to about 1 million tokens.
• Agentic coding focus. The card scores 80.5 on Terminal-Bench 2.1 and 76.5 on SWE-bench Verified. The comparison set is open models from 501B to 975B parameters.
• Local claim. Ricursive says the model fits on consumer hardware such as the NVIDIA DGX Spark, and ships an NVFP4 checkpoint and a Colab notebook for local work. It's also hosted on Together AI, OpenRouter, Fireworks and GMI Cloud.

Read the table, not just the headline. Ripple doesn't win everything. On SWE-bench Verified it trails Reflection Beam (81.6) and Muse Spark Open (78.7). The card calls the posttraining "relatively light" and calls its comparison a "best-effort attempt" because replicating benchmarks was hard. Ricursive's day job is AI chip design; an internal variant of this model powers that platform.

Why an open coding model this size matters for your business

Most strong open models are too big to self-host without a rack. A dense 65B model with a 4-bit checkpoint changes the math. Code that can't leave your network, like client repos or anything under an NDA, can now get a capable agent on one box.

Don't swap your coding stack on a benchmark table. Do this instead:

1. Pick five real tasks from your backlog: a bug fix, a refactor, a test suite, a migration.
2. Run them through Ripple on a hosted endpoint first. It costs little and proves the fit before you buy hardware.
3. Compare against what you use now on pass rate and review time. A fast wrong answer isn't cheap.

Same rule as with the Qwen3.8-27B release: a permissive license is the easy part. Your repo is the only benchmark that counts.

Sources: Ricursive Intelligence, Hugging Face model card.