Ripple-65B claims the US open coding crown for $32M. Its own model card shows how narrow that crown is
A 65B model from a chip-design startup says it beat 501B and 816B models on agentic coding, for $32M all in. Hacker News gave it 330 points. Ricursive Intelligence's Ripple-65B Preview is real, Apache 2.0, and on Hugging Face with several hosting partners. It's also a great lesson in how to read a benchmark claim, because the company published the evidence against its own headline.
The headline: 80.52 on Terminal-Bench 2.1, ahead of Reflection Beam 501B at 80.1, Meta Muse Spark Open 816B at 76.4, and Inkling 975B at 63.8. Yes, a 0.4-point lead, and "US open-weight" scope only. Now open the model card. SWE-bench Verified: Ripple 76.5, Beam 81.6, Muse 78.7, Inkling 77.6, so Ripple is below all three, though some of those rows are Ricursive's in-house runs. SWE-bench Pro v1.0: Ripple 61.42, Beam 66.1. DeepSWE 1.1: Ripple 44.3, Muse 60.0. Ripple does win SWE-bench Pro v2.0 Hard, 79.1 to Beam's 76.5.
Then there's the harness footnote. Under the official Terminus 2 harness in Ricursive's own rerun, Beam falls from 80.1 to 70.8 while Ripple stays at 80.5. That could mean Ripple's lead is bigger than advertised. It could also mean the score belongs to the harness as much as the model. Ricursive says as much: replicating agentic evals "has been challenging". Credit where due for printing all of it.
The cost claim deserves the same squint. $32M is Ricursive's full R&D figure, which is a more honest scope than most. "One-tenth the cost" of the previous leader, though, is "as estimated to the best of our knowledge", and rests on a reported $150M-a-month compute commitment at Reflection. The sharper number is 1.4M rollouts, which it says is roughly 70x fewer than Beam, by designing its own synthetic data instead of buying post-training data. If it holds, that matters more than the crown: a small team with a careful recipe got close to the big kids.
The same day, Cognition shipped SWE-2 Flash, also claiming to outperform all American open-weight models, on different benchmarks: Terminal-Bench 4.0 up from 0.9% to 20% over its Inkling-Small base. It costs one-fifth of SWE-2. But its weights aren't public, citing cyber risk, with a partner program promised "in a few weeks". Two claimants to one crown, one you can download. For a builder that's the only ranking that matters this week: run it on your own task, in your own harness. Ripple: https://www.ricursive.com/ripple Card: https://huggingface.co/ricursiveAI/Ripple-65B-Preview SWE-2 Flash: https://cognition.com/blog/swe-2-flash
← Back to all articles
The headline: 80.52 on Terminal-Bench 2.1, ahead of Reflection Beam 501B at 80.1, Meta Muse Spark Open 816B at 76.4, and Inkling 975B at 63.8. Yes, a 0.4-point lead, and "US open-weight" scope only. Now open the model card. SWE-bench Verified: Ripple 76.5, Beam 81.6, Muse 78.7, Inkling 77.6, so Ripple is below all three, though some of those rows are Ricursive's in-house runs. SWE-bench Pro v1.0: Ripple 61.42, Beam 66.1. DeepSWE 1.1: Ripple 44.3, Muse 60.0. Ripple does win SWE-bench Pro v2.0 Hard, 79.1 to Beam's 76.5.
Then there's the harness footnote. Under the official Terminus 2 harness in Ricursive's own rerun, Beam falls from 80.1 to 70.8 while Ripple stays at 80.5. That could mean Ripple's lead is bigger than advertised. It could also mean the score belongs to the harness as much as the model. Ricursive says as much: replicating agentic evals "has been challenging". Credit where due for printing all of it.
The cost claim deserves the same squint. $32M is Ricursive's full R&D figure, which is a more honest scope than most. "One-tenth the cost" of the previous leader, though, is "as estimated to the best of our knowledge", and rests on a reported $150M-a-month compute commitment at Reflection. The sharper number is 1.4M rollouts, which it says is roughly 70x fewer than Beam, by designing its own synthetic data instead of buying post-training data. If it holds, that matters more than the crown: a small team with a careful recipe got close to the big kids.
The same day, Cognition shipped SWE-2 Flash, also claiming to outperform all American open-weight models, on different benchmarks: Terminal-Bench 4.0 up from 0.9% to 20% over its Inkling-Small base. It costs one-fifth of SWE-2. But its weights aren't public, citing cyber risk, with a partner program promised "in a few weeks". Two claimants to one crown, one you can download. For a builder that's the only ranking that matters this week: run it on your own task, in your own harness. Ripple: https://www.ricursive.com/ripple Card: https://huggingface.co/ricursiveAI/Ripple-65B-Preview SWE-2 Flash: https://cognition.com/blog/swe-2-flash
Comments