Ricursive Intelligence released Ripple-65B Preview, a 65-billion-parameter model, under the Apache 2.0 open license. The company reports 80.52 on the TerminalBench 2.1 coding test. That is ahead of four larger US open models, which have 501 to 975 billion parameters. The closest, Reflection Beam 501B, scored 80.1. Ricursive ran the comparison itself and calls it a best-effort attempt at fairness.
Ricursive says scores fell for every model but Muse Spark Open under the official Terminus 2 test setup at 128K context. So its ranking comes from a different setup, and nothing here is independently verified. The Decoder reports Ripple trails Inkling 2 Small on all three other tests: 79.1% against 86.3% on SWE-Bench Pro v2.0 Hard, 44.3% against 72.1% on DeepSWE, and 71.67% against about 85 on MCP Atlas.
A dense model uses all its parameters for every token. A mixture-of-experts model, the common alternative, uses only some. Ricursive also offers a compact NVFP4 version for local use. This is a Preview with light post-training aimed at agentic coding. With a 0.42-point lead, the buying question is where it runs and who controls it. Test it on your own code, on your own hardware, and read the license.