Ripple-65B Claims 80.5 on Terminal-Bench 2.1 [2026]

Ricursive Intelligence released Ripple-65B Preview, a 65-billion-parameter dense Transformer, under the Apache 2.0 license with weights on Hugging Face at ricursiveAI/Ripple-65B-Preview, and its model card reports a Terminal-Bench 2.1 score of 80.52, which VentureBeat reports as a claimed U.S. open coding lead. The self-reported score edges past Reflection Beam 501B at 80.1 and Meta's Muse Spark Open 816B at 76.4, models 7.7x and 12.6x Ripple's size respectively, though no third-party reproduction of the result has been reported. Ricursive also claims the model runs locally on a MacBook or an NVIDIA DGX Spark, a deployability claim not covered by independent hardware tests, and has not published a date for a final version.

A small U.S. lab just put a 65-billion-parameter model at the top of a table full of far bigger rivals. VentureBeat reports https://venturebeat.com/ai/ex-googlers-new-apache-2-0-model-ripple-65b-preview-claims-the-u-s-open-coding-lead-on-two-major-benchmarks that Ricursive Intelligence released Ripple-65B Preview under the Apache 2.0 license and that the company claims the U.S. open coding lead on Terminal-Bench 2.1 and SWE-bench Pro v2.0 Hard. The model card puts the Terminal-Bench 2.1 score at 80.52, which comparison tables round to 80.5. That number edges past Reflection Beam 501B at 80.1 and clears Meta’s Muse Spark Open 816B at 76.4. We covered how that larger model runs on modest hardware in our look at Intel’s day-0 support for Muse Spark Open https://shattered.io/intel-day-0-support-meta-muse-spark-open-816b-4-gpus-160gb-2026/ , so the comparison lands close to home for anyone tracking open weights this autumn. This analysis separates what the sources confirm from what remains a company claim. The benchmark scores are self-reported on the model card, the “U.S. open coding lead” is Ricursive’s own framing, and the “ex-Googlers” label in VentureBeat’s headline is not clearly established by the other material we reviewed. Those caveats matter, because open coding models are now a purchasing decision for engineering teams, not a curiosity. Our AI and machine learning coverage https://shattered.io/ai-machine-learning/ tracks that shift week by week. Don't miss new tech stories on Google Add Shattered once in the Google app and our stories show up in your news suggestions. Add now https://www.google.com/preferences/source?hl=en&q=https%3A%2F%2Fshattered.io%2F What Ricursive Intelligence Released With Ripple-65B Preview Ripple-65B Preview is a 65-billion-parameter dense Transformer, according to the reports we reviewed. Ricursive Intelligence ships it under Apache 2.0, one of the most permissive licenses in the open-weight market. The weights sit on Hugging Face under the repository name ricursiveAI/Ripple-65B-Preview, and the Hugging Face model page https://huggingface.co/ricursiveAI/Ripple-65B-Preview carries the benchmark tables the company wants developers to read. The word “Preview” in the name is doing real work. Labs use it to signal that post-training, evaluation or safety tuning may still change before a final release. Buyers should treat every score in this article as a snapshot of a preview checkpoint, not a frozen product. Ricursive has not, in the material we found, published a date for a final version. The company’s launch post makes a size-versus-score argument. It says Ripple sets a new state of the art for U.S. open-weight models on agentic coding tasks while beating models that carry several times as many parameters. The same post says the model fits locally on a MacBook or an NVIDIA DGX Spark. That is a company claim about deployability, and independent hardware tests were not part of the retrieved sources. Several outlets picked up the launch within a day. OfficeChai and TokenPost both repeated the headline score, and TokenPost’s report specifically carries the 80.52 figure from the model card. Nobody in that coverage reports a third-party reproduction of the result yet, which is normal for a release this fresh but worth keeping in mind. Terminal-Bench 2.1: How Ripple-65B Preview Stacks Up Terminal-Bench 2.1 is the headline test. It measures how well an AI agent completes real tasks inside a command-line environment, which is much closer to daily engineering work than a multiple-choice coding quiz. The model card compares Ripple against four named rivals, and the spread is wide. | Model | Listed size | Terminal-Bench 2.1 | Gap to Ripple points | Size vs Ripple | |---|---|---|---|---| | Ripple-65B Preview | 65B dense | 80.5 80.52 on card | 0 | 1.0x | | Reflection Beam | 501B | 80.1 | 0.4 | 7.7x | | Meta Muse Spark Open | 816B | 76.4 | 4.1 | 12.6x | | Thinking Machines Inkling | 975B | 63.8 | 16.7 | 15.0x | | NVIDIA Nemotron 3 Ultra | 550B | 56.4 | 24.1 | 8.5x | The gap and size columns are our arithmetic on the published figures. The sizes are headline parameter counts as listed, and the retrieved sources do not give active-parameter counts for the larger models. That caveat matters if any of them use sparse routing, because a sparse model activates only part of its weights per token. Read the table honestly and the story shrinks a little at the top. Ripple leads Beam by 0.4 points, a margin that sits inside the noise of many agentic evaluations. The convincing gaps are against Muse Spark Open https://shattered.io/muse-spark-open-vs-qwen3-8-mcp-atlas-89-2-vs-84-5-2026/ , Inkling and Nemotron 3 Ultra, where Ripple leads by 4.1, 16.7 and 24.1 points. The model card also lists a second Terminal-Bench https://shattered.io/claude-opus-5-5-vs-gpt-6-astra-terminal-bench-2026/ row run with a different harness, and rival scores shift between the two rows. That harness sensitivity is a reminder to rerun the test on your own tasks before you commit. SWE-bench Pro v2.0 Hard and the Other Reported Scores The second benchmark in Ricursive’s claim is SWE-bench Pro v2.0 Hard, where VentureBeat and a second outlet report a score of 79.1 for Ripple. This test asks a model to resolve difficult, real-world software issues, so it rewards sustained reasoning across files rather than quick snippets. A 79.1 would be a strong result for any open model, and the company positions it as the second pillar of the U.S. lead. | Benchmark | Reported Ripple score | What it tests | |---|---|---| | Terminal-Bench 2.1 | 80.52 rounded 80.5 | Agent tasks in a command-line environment | | SWE-bench Pro v2.0 Hard | 79.1 | Hard real-world software issues | | SWE-bench Verified | 76.5 | Resolving human-checked GitHub issues | | MCP Atlas | 71.7 | Tool use through the Model Context Protocol | | IFBench | 79.2 | Following precise instructions | | AIME 2025 | 98.3 | Competition-level mathematics | | GPQA Diamond | 85.2 | Graduate-level science questions | The wider spread shows a model tuned for more than code. A 98.3 on AIME 2025 and an 85.2 on GPQA Diamond suggest strong reasoning, while the 71.7 on MCP Atlas speaks to agentic tool use, which matters for anyone wiring a model into developer tooling. These figures come from the reported results around the launch and, like the others, originate with Ricursive. Ripple does not win every row. The model card’s own comparison table lists at least one rival ahead of it on SWE-bench Verified, and aggregator BenchLM, which tracks a broader field that includes closed models, shows Ripple behind the best verified entries on SWE-bench Pro. The “U.S. open” qualifier in the lead claim is therefore load-bearing. Ripple leads a defined group of open-weight U.S. models on two tests, not the whole market on every test. Why a 65B Dense Model Beating Larger Rivals Matters Open-weight releases through 2026 have chased scale. Moonshot’s Kimi K3.1 Pro went open-weight at 2.8 trillion parameters, as we detailed in our piece on Kimi K3.1 Pro’s open weights https://shattered.io/moonshot-kimi-k3-1-pro-open-weights-2-8-trillion-parameters-2026/ , and Tencent’s Hy4 arrived at 770 billion parameters under Apache 2.0. Against that backdrop, a 65-billion-parameter dense model posting top coding scores flips the usual story. What the parameter gap means for cost Inference cost tracks the weights a model must load and the compute it spends per token. A dense 65B model needs far less memory than an 816B or 975B model, even before any quantization. Ricursive’s claim that Ripple runs locally on a MacBook or DGX Spark would, if it holds in practice, put frontier-style coding help on a developer’s own desk. Independent users will need to confirm the speed and quality at the precision that fits. Dense versus sparse is the open question A dense model uses all its parameters on every token. Many of the biggest open releases use sparse mixture-of-experts layouts, which activate a fraction of their weights. If the larger rivals in the table are sparse, their effective compute per token is lower than their headline size implies, and the 7.7x to 15.0x size gaps overstate the real efficiency advantage. The retrieved sources do not settle this, so we flag it as the most important missing detail in the comparison. Architecture and Context Window: What the vLLM Recipe Shows Serving details leak out through inference recipes before they appear in glossy blog posts. The vLLM recipe for Ripple-65B Preview https://recipes.vllm.ai/ricursiveAI/Ripple-65B-Preview lists 80 layers and support for a 2-million-token context window. Eighty layers is a deep stack for a 65-billion-parameter model, and it fits the pattern of a dense design that spends its parameter budget on depth. The context figure needs care. Ricursive’s own launch post, as quoted in search snippets, cites support for up to 1 million tokens of context, while the vLLM recipe lists 2 million. Where sources conflict, we use the more conservative number for planning, so assume roughly 1 million tokens until the model card or a maintainer clarifies. A long window also does not guarantee good recall across it, and Ricursive has not published long-context retrieval results in the material we found. For teams building coding agents, long context matters because repositories are big. An agent that can hold a large codebase, a test log and a tool transcript in one window makes fewer lossy summarization steps. That is the use case the Terminal-Bench and SWE-bench scores are meant to proxy, and it explains why vendors now publish serving recipes on launch day. Apache 2.0: What the License Means for Developers Apache 2.0 allows commercial use, modification and redistribution, and it includes an express patent grant. Compared with community licenses that cap monthly users or restrict competing services, it removes a whole class of legal review. A startup can fine-tune Ripple, ship it inside a paid product and keep its changes private, provided it keeps the required notices. The license choice also matters for the competitive story. Tencent chose Apache 2.0 for its 770-billion-parameter release, which we covered in our Hy4 report https://shattered.io/tencent-hy4-770b-parameters-apache-2-0-open-source-model-2026/ , and Ricursive follows the same route at a far smaller size. Permissive licensing is becoming the default posture for labs that want adoption over control, especially among teams that cannot send source code to a third-party API. One caution applies to “open.” Open weights are not open training data or open training code, and the reports we reviewed do not describe either. Developers get a model they can run and adapt, but they cannot fully audit how it was built. Security teams should weigh that before they put the model near proprietary repositories. Where to Run It: Hugging Face, Together AI, GMI Cloud and OpenRouter Distribution is already wide for a preview. The weights are on Hugging Face, and sources report hosted access through Together AI, GMI Cloud and OpenRouter. Hosted endpoints let a team test the model against its own prompts in an afternoon, with no GPU purchase. Self-hosting versus hosted APIs Self-hosting through vLLM gives control over data and latency, and it avoids per-token billing once the hardware exists. Hosted routes trade that control for speed of setup. GMI Cloud published its own write-up of agent tests on Ripple-65B Preview https://www.gmicloud.ai/en/blog/we-tested-ripple-65b-preview-with-real-bugs-live-tools-and-planted-traps-here-is-what-held-up , built around real bugs, live tools and planted traps. A host testing its own product is not a neutral reviewer, but the method is the right kind of test for agent behavior. Pricing is the gap. We found no confirmed list price for Ripple in the retrieved sources, and per-token rates differ by provider and change often. Check each host’s pricing page directly before budgeting. Teams that already run local models can follow our Ollama API key and reverse proxy walkthrough https://shattered.io/ollama-api-key-reverse-proxy-docker-12-steps-2026/ to put a gateway in front of whatever they self-host. What Ricursive and Early Observers Are Saying We found no on-record interviews with named engineers, so the voices below come from public launch posts. Ricursive’s official announcement states the core claim in its own words: “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.” That statement comes from Ricursive Intelligence’s launch post on X https://x.com/RicursiveAI/status/2118046765857546733 . It names the four rivals in our first table, and it frames the result as a state-of-the-art claim for U.S. open-weight models specifically. “Very excited to launch Ripple-65B Preview.” Co-founder Azalia Mirhoseini opened her own launch post on X with that line. It carries enthusiasm and no new data, which is what we expect from a founder on release day. “Ok, this is impressive: Ricursive Intelligence just released Ripple-65B Preview, an open-weight AI model that it says outperforms much larger US models on a key agentic coding benchmark.” Technology commentator Chubby wrote that in an independent reaction on X. Note the phrase “that it says,” which keeps the claim attributed to the company. Our reading matches that caution. Reaction posts show attention, and attention is not verification. Reading the Benchmarks Critically Every headline number here is self-reported on the model card. That is standard practice at launch, and it is also where inflated claims creep in. Labs choose the harness, the prompt format, the number of attempts and the comparison set. Change any of those and a 0.4-point lead over Beam can vanish or grow. Three checks separate a durable result from a launch-week spike. First, look for reproduction by a neutral party such as an independent leaderboard or a university group. Second, test contamination risk, meaning whether benchmark tasks leaked into training data. Third, run your own tasks, because a model that tops Terminal-Bench can still stumble on your build system, your monorepo layout or your internal tools. Aggregators will help over the coming weeks. BenchLM already lists Ripple, and community leaderboards typically publish independent runs within days of a notable weights release. If those runs land near 80 on Terminal-Bench 2.1, the lead claim gains weight. If they land ten points lower, the harness or setup deserves a closer look. Confirmed, Claimed and Unproven: A Scorecard | Claim | Status | Basis | |---|---|---| | Apache 2.0 license | Reported consistently | Model listing and multiple outlets | | Weights on Hugging Face | Listed | ricursiveAI/Ripple-65B-Preview | | 80.52 on Terminal-Bench 2.1 | Self-reported | Model card, repeated by TokenPost | | 79.1 on SWE-bench Pro v2.0 Hard | Self-reported | Reported by VentureBeat and another outlet | | U.S. open coding lead | Company claim | Not independently established | | “Ex-Googlers” founders | Unconfirmed | Headline label, not clearly backed by other sources | | 2-million-token context | Conflicting | vLLM recipe says 2M, launch post says up to 1M | | Product pricing | Not found | No confirmed price in the sources | The scorecard shows how much of the story rests on one party. Ricursive wrote the model card, chose the comparisons and made the lead claim. Outlets such as VentureBeat reported those claims accurately, but reporting a claim is different from testing it. We separate the two throughout this article. Competitive Comparison: The Open Coding Field in November 2026 The field is crowded and moves fast. Reflection Beam 501B sits a hair behind Ripple on Terminal-Bench 2.1. Meta’s Muse Spark Open 816B trails by about four points, and we put it head to head with DeepSeek V4.1 Flash and Beam 1.0 in our 1M-context comparison https://shattered.io/muse-spark-open-vs-deepseek-v4-1-flash-vs-beam-1-0-1m-context-2026/ . Inkling and Nemotron 3 Ultra sit far lower on this specific test, though both labs may point to other strengths. Ripple’s pitch is efficiency and permissiveness together. It claims top-tier agentic coding at a fraction of the headline size, under a license with no usage caps. Rivals counter with bigger knowledge bases, broader multilingual coverage or ecosystem support. Our open-weight spec sheet https://shattered.io/open-weight-ai-models-side-by-side-the-spec-sheet/ lays out those trade-offs for the wider group. Closed models remain the benchmark ceiling. Aggregators that include proprietary systems show them ahead of Ripple on several coding tests, and developers who pay for managed assistants still get polish that open weights lack. Our Claude Code, Cursor and Copilot pricing comparison https://shattered.io/claude-code-vs-cursor-vs-github-copilot-pricing-limits-2026/ shows what those subscriptions cost, which is the number a self-hosted Ripple deployment has to beat. Market Impact for Coding Agents and Tool Vendors A strong, small, permissive coding model changes bargaining power. Agent startups that pay a closed vendor per token now have a credible self-hosted alternative to price against. Even teams that never switch gain leverage in contract talks, and that pressure tends to push API prices down across the market. Hosting providers benefit directly. Together AI, GMI Cloud and OpenRouter added Ripple quickly, and each gets inference demand from developers who want to test it. Chip and server vendors benefit too, since a 65B dense model fits a much wider range of machines than an 816B or 975B one. Enterprises with data-residency rules gain a path to run agentic coding inside their own network. The pressure lands hardest on mid-tier closed models that sell on coding skill alone. If an Apache 2.0 model matches them on public tests, they need to justify their price through tooling, support, safety guarantees or integration depth. Vendors with those advantages will hold up, and vendors without them will feel the squeeze first. Historical Context: Open Weights Climb the Coding Ladder Open models spent years a generation behind closed ones on code. That gap narrowed as more labs shipped large open-weight systems, and by 2026 the argument moved from whether open models could code to which ones could run agents. Terminal-Bench and the SWE-bench family became the scoreboards because they test work, not trivia. The U.S. framing in Ricursive’s claim reflects a policy and market mood. The qualifier signals that Ricursive wants to be read as a domestic open-weight option. That is why the “U.S. open” qualifier appears in the headline, and why we treat it as a positioning statement more than a technical one. Size trends also changed. The 2026 releases we covered ran from 280 billion parameters in Baidu’s ERNIE 5.5 Flash Base to 2.8 trillion in Kimi K3.1 Pro. Ripple’s 65 billion sits well below that range. If independent testing confirms its scores, it will bolster the argument that training recipe and data quality can beat raw scale on agentic work. Five Predictions for Ripple-65B Preview and Open Coding Models These are our forecasts, not reported facts. - Independent runs will land lower than the model card. Self-reported numbers usually slip a few points under third-party harnesses, and we expect Ripple to stay near the top of the open group anyway. - The Beam gap will blur. A 0.4-point margin will not survive repeated runs, so the open lead may end up shared between Ripple and Beam. - Quantized and fine-tuned variants will multiply. Apache 2.0 and a 65B size invite community builds within weeks. - Rivals will answer with smaller models. Labs with 500B-plus systems face pressure to show efficient versions, since Ripple turned size into a liability in the headline table. - A final release will follow the preview. The name signals more work, and we expect Ricursive to publish updated weights and fuller evaluation detail. What Developers Should Do This Week Start with a bounded test. Pull a hosted endpoint from one of the reported providers and run 20 to 30 real tasks from your own backlog, scoring each one pass or fail. Compare the results against the model you use today. A small, honest internal benchmark beats any public leaderboard for your decision. If the results look good, test the self-hosted path next. Check memory needs at the precision you plan to use, confirm throughput under your real concurrency, and look at the context length you actually get. Hold off on production traffic until independent evaluations appear. The takeaway: Ripple-65B Preview deserves a serious trial, and it has not earned blind trust yet. FAQ: Ripple-65B Preview What is Ripple-65B Preview? Ripple-65B Preview is a 65-billion-parameter dense Transformer coding model from Ricursive Intelligence. It ships under the Apache 2.0 license, and its weights are listed on Hugging Face as ricursiveAI/Ripple-65B-Preview. How well does it score on Terminal-Bench 2.1? The model card reports 80.52, which comparison tables round to 80.5. Reflection Beam 501B is listed 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. Is it really the top U.S. open coding model? That is Ricursive’s claim, based on its own reported benchmark results. The sources we reviewed do not independently establish it, so treat it as a company claim until neutral testing arrives. Did former Google engineers build it? VentureBeat’s headline uses that label. The other material we reviewed does not clearly establish it, and we do not repeat it as fact. Where can I run Ripple-65B Preview? You can download the weights from Hugging Face or use hosted access that sources report through Together AI, GMI Cloud and OpenRouter. A vLLM recipe exists for self-hosting. How big is the context window? Sources disagree. The vLLM recipe lists 2 million tokens, while the launch post cites up to 1 million. Plan around the lower figure until Ricursive clarifies. How much does it cost? We found no confirmed price. The weights are free under Apache 2.0, but hosted providers set their own per-token rates, so check each provider’s pricing page.