Ricursive releases Ripple-65B, claims open-source agentic coding SOTA on a $32M budget

Ricursive Intelligence (RicursiveAI) has released its first large model, Ripple-65B Preview: a 65B dense LLM pretrained from scratch on 12 trillion tokens, supporting up to 1M context. The company claims it sets a new SOTA for US open-weight models on agentic coding, topping TerminalBench 2.1 and significantly outperforming much larger models (including Muse Spark), with total R&D/compute investment of around $32 million — reportedly just one-tenth the cost of the previous open-source agentic coding SOTA. Model weights are now available on HuggingFace and OpenRouter.

Confirmed

• Ripple-65B Preview was released by Ricursive Intelligence: 65B dense, pretrained from scratch, 12 trillion tokens, up to 1M context
• Official claim: achieves agentic coding SOTA among US open-weight models on TerminalBench 2.1, surpassing larger models (e.g., Muse Spark)
• The company states total investment was $32 million, roughly one-tenth the cost of the previous open-source SOTA
• The model is available on HuggingFace and OpenRouter
• Stanford professor and former core member of Google's chip-design AI team Anna Goldie reshared the team's announcement

Not Yet Confirmed

• @teortaxesTex questioned the "comparison methodology," noting the benchmark comparisons underlying the claimed SOTA and cost advantage remain disputed, with official details pending further verification

Why It Matters

• If the cost and performance claims hold, it means the open-source agentic coding frontier can be reached on a far smaller budget, shaking up the cost-effectiveness landscape of open-source coding models
• Ricursive is a chip-design AI company, and this is its first LLM — the cross-domain entry itself is noteworthy, and endorsements from academics like Anna Goldie add to the buzz

2026-11-05 ~ 2026-11-05 · 13 related posts

• [source] Ricursive releases Ripple-65B Preview, claiming US open-weight SOTA on agentic coding — DZhang50 · 2026-11-05
• [source] Ricursive Intelligence launches Ripple-65B, claiming open-weight agentic coding SOTA — xyzrt · 2026-11-05
• Ripple claims $32M compute for open-weight coding SOTA; skeptics question the math — teortaxesTex · 2026-11-05
• Ricursive AI launches 65B dense model from scratch, claims open-weight agentic coding SOTA — simonguozirui · 2026-11-05
• Dark horse Ripple-65B Preview claims US open-weight SOTA for agentic coding, runs on a MacBook Pro — TheZachMueller · 2026-11-05
• ricursiveAI releases Ripple-65B-Preview weights on Hugging Face — coder543 · 2026-11-05
• Ripple-65B: 65B dense model claims open-weight agentic coding SOTA, beats 500B+ MoEs — simonguozirui · 2026-11-05
• Ripple-65B released: new SOTA for U.S. open-weight models in agentic coding — DZhang50 · 2026-11-05
• 5-person team trains Ripple-65B from scratch, sets US open-weight SOTA on Terminal-Bench 2.1 — annadgoldie · 2026-11-05
• Ricursive releases Ripple-65B open weights: $32M model hits US open-weight SOTA on TerminalBench 2.1 — DZhang50 · 2026-11-05
• [source] RicursiveAI releases Ripple-65B Preview, first open model trained on NVIDIA B300 with NVFP4 checkpoint — Azaliamirh · 2026-11-05

2 near-duplicate retellings: annadgoldie · annadgoldie