Ricursive Intelligence releases Ripple-65B Preview, a 65B dense open-weight model under Apache 2.0

Ricursive Intelligence, a firm known for AI chip design, put Ripple-65B Preview on Hugging Face under the Apache 2.0 licence. It is a dense 65B-parameter Transformer pretrained on 12 trillion tokens and tuned for Nvidia hardware. The company reports 80.52 on TerminalBench 2.1, which it calls the best result among open US models, and says development and training cost $32 million. It also reports lower scores than 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 about 85%). The coverage links to no model card or other primary source, so none of these figures has been checked independently.

Developers and enterprises that want to self-host a coding-capable model with no licence restrictions get another option. Apache 2.0 allows commercial use and modification, and a dense 65B model is simpler to deploy than larger or more complex designs. The practical case is narrow, though. The model trails Inkling 2 Small on three of the four benchmarks reported, and by a wide margin on DeepSWE, so teams doing general software engineering work have little reason to switch yet. Ricursive is not a leading lab, and this is labelled a preview, so the release signals little about the wider market.

"Best open US model" rests on one benchmark, reported by the company itself, with harnesses it chose, and the qualifiers "open" and "US" shrink the field it is compared against. The $32 million figure is self-reported with no breakdown, so it cannot be compared with other labs' costs. The headline "strong agentic coding results" sits awkwardly with the company's own numbers, which show it losing badly on several coding and tool-use tests.

1. 01Independent runs of TerminalBench 2.1, SWE-Bench and DeepSWE on the released weights, and whether they land near the company's 80.52 and the lower figures it admits.
2. 02Hugging Face download counts, community fine-tunes, and any reports of teams deploying Ripple-65B in production over the next month, set against Inkling 2 Small.
3. 03Whether the full model card, training data details and a final non-preview release follow, and whether the $32 million cost claim gets any supporting breakdown.