Ricursive releases Ripple-65B Preview, an open coding model

Ricursive Intelligence released Ripple-65B Preview on November 4, an open-weight coding model that the lab says sets a new U.S. open-weight high on Terminal-Bench 2.1. Ripple-65B Preview scores 80.5 on that agentic coding test, ahead of models up to 15 times its size, by the company's own measurement.

Most open models that compete on coding are hundreds of billions of parameters. This one has 65 billion, is dense (every parameter is used for every token), and the company says it fits on a single NVIDIA DGX Spark desktop machine. That makes it a candidate for teams that want a strong coding agent they can host themselves.

What Ricursive released

Ricursive is a lab that works on AI chip design. Ripple is a general-purpose variant of the model behind that platform. The model card on Hugging Face lists these details:

• 65B parameters, dense Transformer, 80 layers
• Pretrained on 12 trillion tokens, mostly English and code
• Context of up to 1 million tokens
• Apache 2.0 license
• BF16 weights, plus an NVFP4 checkpoint, a base model and a DFlash2 variant

Ricursive says it spent about $32 million on the whole effort, from data mixing to the final training run, and used about 1.4 million reinforcement learning rollouts, roughly 70 times fewer than Reflection Beam.

Benchmark scores, and who reported them

On Terminal-Bench 2.1, which tests agents working in a terminal, Ricursive reports:

The card also lists 76.5 on SWE-bench Verified, 85.2 on GPQA Diamond and 95.0 on AIME 2026. Every score is self-reported, and some are in-house evaluations. The authors add that agentic scores are hard to reproduce: results were generally lower under the official Terminus 2 harness at 128K context with summarization. Treat the headline number as a claim until outside testers run it.

Price and where to run it

The weights are on Hugging Face. As of November 5, 2026, OpenRouter lists the model through GMI Cloud at $0.20 per million input tokens and $0.60 per million output tokens. Ricursive's page also names Together AI and GMI Cloud, and says Fireworks would follow. The model card notes the repository uses custom code, so loading it may need extra steps.

What we don't know yet

This is a preview with light post-training aimed at agentic coding. The training data is mostly English. No independent benchmark runs exist yet, and the company hasn't said when a final version will ship. The next thing to watch is whether Artificial Analysis or the Terminal-Bench maintainers publish their own numbers for it.