US Startup Shocks Open Model Race

🔥 What happened Ricursive Intelligence, a chip-design startup, dropped Ripple-65B Preview – an open-weights model it claims beats every US rival on Terminal-Bench 2.1 and SWE-bench Pro v2.0 Hard. The 65B model is live on Hugging Face and via APIs at $0.20 per million input tokens. 💡 Why it matters Ripple reportedly outperforms models up to 15x its size – and cost just $32M to train, a fraction of the hundreds of millions burned by frontier labs. But the benchmark numbers contradict themselves internally, and Meta crushes it on DeepSWE, 60 to 44. ⚡ Our take A 65B model dunking on 975B rivals would be seismic – if the numbers hold. The mismatches between launch graphic and model card smell like benchmark shopping. Test it yourself before rewiring your pipelines.