--- title: Ricursive AI releases Ripple-65B Preview, a 65B dense... description: Ricursive AI released Ripple-65B Preview, a 65-billion-parameter dense language model pretrained from scratch. The company claims it sets a new state of the... canonical: https://daily.dev/posts/ricursive-ai-releases-ripple-65b-preview-a-65b-dense-llm-pretrained-from-scratch-yxwfa9jlo twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Ricursive AI releases Ripple-65B Preview, a 65B dense LLM pretrained from scratch | daily.dev og:description: Ricursive AI released Ripple-65B Preview, a 65-billion-parameter dense language model pretrained from scratch. The company claims it sets a new state of the... og:url: https://daily.dev/posts/ricursive-ai-releases-ripple-65b-preview-a-65b-dense-llm-pretrained-from-scratch-yxwfa9jlo og:image: https://api.daily.dev/og/posts/YXWfa9JLO.png og:image:alt: Ricursive AI releases Ripple-65B Preview, a 65B dense LLM pretrained from scratch og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Ricursive AI releases Ripple-65B Preview, a 65B dense LLM pretrained from scratch **[Collections](https://daily.dev/sources/collections)** · 2 min read · 1 upvotes · 0 comments ## Summary Ricursive AI released Ripple-65B Preview, a 65-billion-parameter dense language model pretrained from scratch. The company claims it sets a new state of the art, but the announcement cuts off before naming the category or benchmarks. The release was reshared by Clement Delangue and iScienceLuvr. Training a dense 65B model from scratch means Ricursive AI bore the full pretraining cost itself rather than fine-tuning an existing base model. The 'Preview' label suggests a fuller version is planned, but Ricursive AI has not said when. ## Content Ricursive Intelligence, a chip design lab, has released Ripple-65B Preview, a 65 billion parameter dense model pretrained from scratch. It scores 80.52 on Terminal-Bench 2.1, the top self-reported result among US open-weight models. ## What it is Ripple is a dense Transformer, not a Mixture-of-Experts model. It was pretrained on 12 trillion tokens and supports up to 1M tokens of context. The weights ship under Apache 2.0 as BF16, NVFP4 and base checkpoints, with an instruct version also available. Ricursive says it holds 40% MFU without custom kernels. You can get it on Hugging Face, Together AI, GMI Cloud and OpenRouter. ## The numbers On Terminal-Bench 2.1, Ripple beats much larger MoE models. Reflection Beam 501B scores 80.1 and NVIDIA Nemotron 3 Ultra 550B scores 56.4. Rivals in this comparison have 501B to 975B total parameters, so a 65B dense model on top is the surprising part. It got there with about 1.4M RL rollouts. Ricursive says that is roughly 70x fewer than the 100M+ reported for Beam. Cost is the other headline. Ricursive puts the full R&D bill at $32M, about a tenth of what the previous open-weight leader reportedly cost. That figure covers everything: pretraining ladder development, data mixing, evals, debugging, synthetic data generation, posttraining recipe development, and the final pretraining and posttraining run. @swyx called it "the avengers of open model SOTA." ## Where it falls short The claim is limited to US open-weight models. Several Chinese open models still score higher on Terminal-Bench 2.1, including DeepSeek Flash 4.1 at 90.6 and Xiaomi's MiMo 2.6 Pro at 89.9. Ripple also trails larger models on other benchmarks: SciCode (39.4), DeepSWE 1.1 (44.3) and MCP Atlas (71.67). All the scores are self-reported. ## Reaction @natolambert described it as the smartest Western open-weight model by a margin, despite being much smaller than Muse Spark open or Mistral Large. @eliebakouch called it a very nice open-weight release with a modern dense model and strong benchmark scores. Clement Delangue, Thom Wolf and others amplified the announcement. A chip design lab topping the US open-weight board on a $32M budget is an odd result. I'd want independent evals before reading too much into it, but the efficiency claims alone make this worth watching. ## Questions this post answers ### How much did it cost to train the Ripple-65B open-weight model? Ripple-65B Preview, a 65-billion-parameter dense LLM from RicursiveAI, cost $32M to develop and train. That figure is the full R&D bill, including pretraining ladder development, data mixing, evals, debugging, synthetic data generation, posttraining recipe development, and the final hero pretraining and posttraining run, not just the final training run alone. _Comparing training costs across open-weight models gets easier when daily.dev surfaces releases like this one._ ### Which open-weight LLM claims state of the art on agentic coding benchmarks at low training cost? Ripple-65B Preview, released by RicursiveAI, is claimed to set a new state of the art for open-weight models on agentic coding. Its developers say it cost about one-tenth the price of the previous open-weight SOTA model on agentic coding, at roughly $32M total R&D spend. _Developers picking an open-weight model for coding agents can track claims like this through daily.dev._ ## Community take How the wider developer community reacted, aggregated from 1 discussion and 93 comments across hackernews (as of 2026-11-06). **TL;DR:** Most of the discussion is skeptical of the 'SOTA' and 'frontier' framing, arguing the benchmark comparison is narrowly cherry-picked to exclude stronger non-US and MoE models, though some see real value in a dense model sized for high-memory local/air-gapped use and in the US-only compliance niche. **Sentiment:** 15% positive · 30% mixed · 55% skeptical **The case for** - Some see genuine utility for air-gapped or government/enterprise deployments that can't use Chinese models. - A dense ~65B model is seen as a good fit for high-memory Mac Studio setups and similar local hardware. - The low $32M training cost and sample-efficient RL rollouts are seen as a promising efficiency trend. - Apache-2.0 licensing is noted as a genuine advantage over some other 'open' models with usage restrictions. **The pushback** - Many argue the comparison chart cherry-picks weaker/older models (e.g. Nemotron 3 Ultra instead of 3.5 Super, Inkling instead of Beam 1.0) and omits stronger options like Qwen. - The 'SOTA'/'frontier' claim is criticized as being narrowed to an artificially small category (US, open-weight, size) to manufacture a win. - Terminal Bench 2.1 is called an outdated benchmark, with some newer benchmarks (TerminalBench 4.0) considered more relevant. - Dense 65B architecture is seen as impractical for consumer hardware compared to MoE models of similar or larger size. - Some question whether the 'U.S.' framing is meaningful at all beyond marketing/compliance positioning. **By community** - hackernews (skeptical): Extensive debate over whether 'SOTA'/'frontier' claims are meaningful given a narrowly defined comparison category, alongside complaints about cherry-picked benchmarks and omitted competitors. **Hottest debate:** Whether narrowing the comparison to 'US open-weight' models makes the SOTA/frontier claim meaningless marketing versus a legitimately useful distinction for compliance-driven buyers. **Open questions** - How would the model actually compare against Qwen and other top open-weight models on the same benchmarks? - How practical is running this dense model locally compared to MoE alternatives of similar size? - Is the benchmark performance 'benchmaxxed' rather than reflecting real-world agentic coding ability? **Highlights** > Category narrowing to this degree really has diluted the term "SOTA" quite a lot You mean it's state of the art for this subset of benchmarks you're targeting? No? Well surely it's state of the art for models of a certain size category? Huh, what about models that are available for local inference? O, state of the art for open models of a certain size made in the US specifically. Uh. Neat! > — [advael on hackernews · 8 comments](https://news.ycombinator.com/item?id=50370633) > What's funnier is that after cherry picking models for comparison, they further cherry picked Inking for size comparison instead of better and smaller Beam 1.0 The bar chart also shows Nemotron 3 Ultra instead of smaller and better Nemotron 3.5 Super. > — [karmakaze on hackernews](https://news.ycombinator.com/item?id=50370722) > It's great to see another competitive open weight model. Congratulations! But as is, it's probably hard to become popular amongst hobbyists and local model enthusiasts like Qwen's 27B class models or MOE models because it's a 65B dense model and inference will be prohibitively slow on consumer grade hardware, and seems like the goal/mission of the company behind this model is to automate chip design. So if I have to assume and connect the dots, this is for the companies who want to use AI in air-gapped environments for chip design so that they protect their IPs (not leaking to third party AI labs/hostile parties) and can afford lots of server grade inference hardware for a quick unlock with the big dense model. Neat! > — [stewartpark on hackernews · 4 comments](https://news.ycombinator.com/item?id=50371154) > I mean Qwen 3.8 27B hits 73% on TB2.1 and 42.2% on DeepSWE.... > — [ethan-boneh on hackernews](https://news.ycombinator.com/item?id=50370612) > So the model is "meh" at best, but the cost of the training is interesting. So we can extrapolate that in 3 years the cost will be below $5m for a model like this. It's very well within the capabilities of any reasonable organization, maybe even crowd-funded. > — [cyberax on hackernews](https://news.ycombinator.com/item?id=50371956) **Source threads** - [hackernews](https://news.ycombinator.com/item?id=50369925) · 243 points · 93 comments ## Similar posts on daily.dev - [It's Swapping Season](https://daily.dev/posts/it-s-swapping-season-uhw8ifpqf) · Latent Space · 0 upvotes · 0 comments - [Building AIE CODE: The top Coding Agent Conference](https://daily.dev/posts/building-aie-code-the-top-coding-agent-conference-ovcg1zp0l) · Latent Space · 0 upvotes · 0 comments --- Tags: [#open-source](https://daily.dev/tags/open-source), [#llm](https://daily.dev/tags/llm) [View this post on daily.dev](https://daily.dev/posts/ricursive-ai-releases-ripple-65b-preview-a-65b-dense-llm-pretrained-from-scratch-yxwfa9jlo) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Ricursive AI releases Ripple-65B Preview, a 65B dense LLM pretrained from scratch","url":"https://daily.dev/posts/ricursive-ai-releases-ripple-65b-preview-a-65b-dense-llm-pretrained-from-scratch-yxwfa9jlo","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/ricursive-ai-releases-ripple-65b-preview-a-65b-dense-llm-pretrained-from-scratch-yxwfa9jlo"},"datePublished":"2026-11-04T19:45:21.637Z","dateModified":"2026-11-06T15:01:35.942Z","description":"Ricursive AI released Ripple-65B Preview, a 65-billion-parameter dense language model pretrained from scratch. The company claims it sets a new state of the...","isAccessibleForFree":true,"articleSection":"Collections","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Collections","logo":"https://media.daily.dev/image/upload/s--fk_6ycEi--/f_auto,q_auto/v1780996001/logos/collections?_a=BAMAMiWQ0","url":"https://daily.dev/sources/collections"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/ricursive-ai-releases-ripple-65b-preview-a-65b-dense-llm-pretrained-from-scratch-yxwfa9jlo","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"open-source,llm","timeRequired":"PT2M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Collections","item":"https://daily.dev/sources/collections"},{"@type":"ListItem","position":3,"name":"Ricursive AI releases Ripple-65B Preview, a 65B dense LLM pretrained from scratch"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/ricursive-ai-releases-ripple-65b-preview-a-65b-dense-llm-pretrained-from-scratch-yxwfa9jlo#faq","mainEntity":[{"@type":"Question","name":"How much did it cost to train the Ripple-65B open-weight model?","acceptedAnswer":{"@type":"Answer","text":"Ripple-65B Preview, a 65-billion-parameter dense LLM from RicursiveAI, cost $32M to develop and train. That figure is the full R\u0026D bill, including pretraining ladder development, data mixing, evals, debugging, synthetic data generation, posttraining recipe development, and the final hero pretraining and posttraining run, not just the final training run alone. Comparing training costs across open-weight models gets easier when daily.dev surfaces releases like this one."}},{"@type":"Question","name":"Which open-weight LLM claims state of the art on agentic coding benchmarks at low training cost?","acceptedAnswer":{"@type":"Answer","text":"Ripple-65B Preview, released by RicursiveAI, is claimed to set a new state of the art for open-weight models on agentic coding. Its developers say it cost about one-tenth the price of the previous open-weight SOTA model on agentic coding, at roughly $32M total R\u0026D spend. Developers picking an open-weight model for coding agents can track claims like this through daily.dev."}}]} ```