Ricursive Intelligence has released the weights of Ripple-65B Preview, a 65-billion-parameter dense model, under an Apache 2.0 license on Hugging Face. Its model card reports a Terminal-Bench 2.1 score of 80.5, against 80.1 for Reflection Beam, a 501-billion-parameter open-weight model, in the card’s self-reported row. That 0.4-point lead comes from a model with about 7.7 times fewer total parameters, by our arithmetic.
The lead is narrow and mostly confined to coding. Ripple has the top score in 3 of the 22 benchmark rows in its own comparison table, by our count; one of the four larger US open-weight models it lists scores higher in the other 19.
Business Pill · Open Weights Explained
An evergreen Business Engineer explainer on open weights: rent a model through an API or run it yourself. It teaches the general idea only and does not discuss Ripple.
Terminal-Bench 2.1 scores from Ricursive Intelligence’s model card (4 November 2026), self-reported row against the Terminus 2 harness row. Rivals’ Terminus 2 scores are Ricursive’s in-house runs, not independently verified. Chart: FourWeekMBA.
The key insight: Ricursive’s Ripple-65B Preview reports 80.5 on Terminal-Bench 2.1 against 80.1 for the 501-billion-parameter Reflection Beam, a model about 7.7 times its total size by our arithmetic; by our count it tops 3 of the 22 benchmark rows in its own table, and it uses more active parameters per token than every model it is compared with.
The model card describes Ripple-65B Preview as “a 65B parameter dense Transformer model pretrained on 12 trillion tokens with support for up to 1M context length”. The Hugging Face repository, created on 4 November, holds 16 weight files; their metadata puts the parameter count at 64,782,689,360.
Ricursive also posted a base model, an NVFP4 version for local hardware and a DFlash2 variant. Its launch blog says the model is “a general-purpose variant of the proprietary model that powers our AI chip design platform”.
Terminal-Bench 2.1 scores in the self-reported row of Ricursive Intelligence’s Ripple-65B Preview model card (4 November 2026), with total parameters. The scores are the company’s own table, not independently verified; the 7.7x size ratio and the 3-of-22 count are our arithmetic.
What Was Posted on X
Today we release Ripple-65B Preview, our 65B dense LLM we pretrained from scratch.
Ripple sets a new state-of-the-art for U.S. open-weight models in agentic coding tasks, surpassing significantly larger models including Muse Spark Open-816B, Beam-501B, Inkling-975B, and Nemotron 3 Ultra-550B.
— Ricursive Intelligence (@RicursiveAI) Nov 04 18:24
Ricursive Intelligence (@RicursiveAI) announced Ripple-65B Preview at 18:24 UTC on 4 November 2026 as its 65B dense LLM pretrained from scratch. The company’s post claims Ripple sets a new state of the art for U.S. open-weight models in agentic coding tasks, surpassing larger models it names, including Muse Spark Open-816B, Beam-501B, Inkling-975B and Nemotron 3 Ultra-550B.
THIS IS SO EXCITING
A new open-source American LLM pretrained from scratch!!!!
I know @DZhang50 has been working overtime to get this out! 🙂
— Tanishq Mathew Abraham, Ph.D. (@iscienceluvr) Nov 04 19:31
Tanishq Mathew Abraham (@iScienceLuvr) quoted that launch post at 19:31 UTC on 4 November 2026, calling Ripple “a new open-source American LLM pretrained from scratch”.
> we developed and trained Ripple-65B Preview for only $32M, which is one-tenth the cost of the previous open-weight SOTA on agentic coding¹
¹ “Note that $32M is our full R&D cost starting from scratch, including pretraining ladder development, data mixing, evals, debugging, synthetic data generation, posttraining recipe development, and the final hero pretraining and posttraining run.”
swyx (@swyx) posted at 20:11 UTC on 4 November 2026, quoting Ricursive as saying it “developed and trained Ripple-65B Preview for only $32M, which is one-tenth the cost of the previous open-weight SOTA on agentic coding”, with a quoted footnote saying the $32M is the full R&D cost starting from scratch, including pretraining ladder development, data mixing, evals and the final pretraining and posttraining run. The one-tenth comparison is Ricursive’s claim. He called it “like the avengers of open model SOTA”.
The Benchmark Claim
On Terminal-Bench 2.1, the card lists Ripple at 80.5, Reflection Beam 501B at 80.1, Meta Muse Spark Open 816B at 76.4, Thinking Machines Inkling 975B at 63.8 and NVIDIA Nemotron 3 Ultra 550B at 56.4. Ripple also leads on SWE-bench Pro v2.0 Hard, at 79.1 against 76.5 for Reflection Beam.
Elsewhere it trails. Reflection Beam leads on SWE-bench Verified (81.6 vs 76.5) and GPQA Diamond (90.6 vs 85.2); Muse Spark Open leads on DeepSWE 1.1 (60.0 vs 44.3) and MCP Atlas (89.2 vs 71.67).
The comparison mixes sources. The card says: “Our table below shows official model card scores where available, and we use our own in-house eval flow for missing evals”. It adds: “Replicating benchmark scores has been challenging, particularly for modern agentic evals”.
Size, Cost and the Active-Parameter Catch
Total parameters favour Ripple: Inkling’s 975 billion is 15 times its 65 billion, by our arithmetic. Active parameters do not. Every model Ripple is compared with is a mixture-of-experts design that uses fewer parameters per token: Reflection Beam lists 23 billion active, so Ripple runs about 2.8 times as many, also our arithmetic.
On cost, the blog says Ricursive “developed and trained Ripple-65B Preview for only $32M”, which it calls “one-tenth the cost of the previous open-weight SOTA on agentic coding” and describes as estimated to the best of its knowledge. It says the $32 million is its full R&D cost from scratch.
As we read the card, the result is a coding result: Ripple’s three top scores are all agentic-coding rows, while larger models lead on knowledge, math, safety and most tool-use rows.
As we read the size claim, “smaller” depends on the measure: Ripple has far fewer total parameters than its rivals but runs all 65 billion on every token, while Reflection Beam lists 23 billion active.
Ricursive Intelligence, Ripple-65B Preview model card
“Replicating benchmark scores has been challenging, particularly for modern agentic evals, due to the number of factors which can cause eval scores to shift”
Three Implications
FOR DEVELOPERS The card says the 65B model can fit on consumer hardware such as the NVIDIA DGX Spark, and Ricursive provides an NVFP4 checkpoint.
FOR BENCHMARK READERS Several rival scores in the card are Ricursive’s in-house runs; under the Terminus 2 harness the card lists Reflection Beam at 70.8 against 80.1 self-reported.
FOR TRAINING BUDGETS Ricursive says its full R&D cost was $32M, which it calls one-tenth the cost of the previous open-weight leader on agentic coding, by its own estimate.
The framework says: “Timing, execution, and competitive dynamics determine whether the strategy actually works.”
As we read it, Ricursive is splitting its layers: the blog calls Ripple a general-purpose variant of the proprietary model that powers its AI chip design platform, so the open weights sit next to a closed product. The framework does not discuss Ricursive.
Ripple’s scores are Ricursive’s own; we found no independent evaluation in what we read. Several rivals’ scores in the table are from Ricursive’s in-house runs, marked in italics on the card, not from those labs. The $32 million and the one-tenth comparison are Ricursive’s estimates. The release is labelled a Preview, and the card says its post-training was relatively light.
This Week In The AI Supercycle
Anthropic’s Reality Check, The Capital Squeeze, & The New World Order
This week’s issue: for two years the AI build-out was priced as a growthstory. This week it started being priced as a credit story.
Ricursive’s Ripple-65B Preview reports 80.5 on Terminal-Bench 2.1 against 80.1 for the 501-billion-parameter Reflection Beam, with weights out under Apache 2.0. By our count it leads 3 of the 22 rows in its own table, and it uses more active parameters per token than every model it is compared with.
A note on sourcing. Model details, benchmark scores and quotations are from Ricursive Intelligence’s Hugging Face model card and launch blog, read on 4 November 2026; repository details are from Hugging Face. Benchmark scores are the company’s own and some rival scores are its in-house runs. Ratios and counts are our arithmetic.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
