Ricursive Intelligence releases Ripple-65B Preview agentic coding model

ricursive.com3 min read

Ricursive Intelligence has released Ripple-65B Preview, a 65-billion parameter open-weight language model optimized for agentic coding and AI chip design. Trained from scratch for $32 million using a compute-efficient dense Transformer architecture and custom reinforcement learning recipes, the model achieved a U.S. open-weight state-of-the-art score of 80.52 on TerminalBench 2.1. Open article →

Notable comments

Overall reception is positive about another open-weight release, but high-signal criticism centers on narrow/misleading SOTA benchmarking (especially vs MoE and vs actual open leaders), plus practical runability concerns for hobby hardware and missing EDA/chip-design coverage.

HN · namjh Points out the benchmark chart is misleading because the comparisons mix dense vs MoE models; suggests adjusting with an “effective param size” heuristic (e.g., effective param size ≈ sqrt/weighted by active experts) to make the score-vs-parameter comparison fair.

HN · 2001zhaozhao Uses concrete reference points (Qwen 3.8 Flash Next scoring 84.3 vs Ripple-65B 80.5, and Qwen 3.8-27B at 73.0) and argues Ripple’s competitiveness depends on how much its score was tuned/benchmaxxed.

HN · stewartpark Notes the 65B dense model is likely too slow for consumer/local inference (unlike 27B dense or MoE), inferring that the product goal is automation of chip design rather than hobbyist deployment; questions whether the release aligns with local-user expectations.

More on this topic