Ripple-65B is an Apache 2.0 open model that its makers say scores 80.52 on TerminalBench 2.1 and fits on a DGX Spark

A permissive license and a model size one box can run is more useful to most client work than another leaderboard record.

What happened

Ricursive Intelligence, an AI chip design company, released Ripple-65B Preview, a 65 billion parameter dense Transformer under the Apache 2.0 license. The company says it was pretrained on 12 trillion tokens, supports up to 1 million tokens of context, and scores 80.52 on the TerminalBench 2.1 coding benchmark, which it calls the best result among US open-weight models. Ricursive puts its full research and development cost at $32 million.

The model card says that at 65B parameters it can fit on consumer hardware such as the NVIDIA DGX Spark, and reaches 250 tokens per second on two B200 GPUs at batch size 8 before quantization. Ricursive says it is available on Hugging Face, Together AI, GMI Cloud, OpenRouter and Fireworks.

It does not lead everywhere. The Decoder's comparison shows it behind Thinking Machines' Inkling 2 Small on SWE-Bench Pro v2.0 Hard (79.1% vs 86.3%), DeepSWE v1.1 (44.3% vs 72.1%) and MCP Atlas (71.67% vs 85.4%), though The Decoder notes Ripple is likely easier to run on your own hardware.

My take

Two details matter for operators more than the coding score.

1. Apache 2.0 means you can host it for a client with strict data rules without licensing headaches. That is the real unlock for clinics, law offices and finance teams that cannot send records to a third party API.
2. MCP Atlas measures tool use across connected services. A 14 point gap there means I would not hand Ripple a multi-tool CRM agent yet. For narrow jobs like summarizing tickets or drafting replies from a fixed template, it is worth testing.

The order I would test it: run it through OpenRouter or Together on 30 to 50 real tasks from the client's workflow, compare against the model they use now, and only buy hardware if it wins on that test and data rules require it. Benchmarks are the makers' numbers, and your workflow is the only test that counts.