{"slug": "ripple-65b-benchmarks-a-dense-model-that-punches-above-its-weight", "title": "Ripple-65B Benchmarks: A Dense Model That Punches Above Its Weight", "summary": "Recursive Intelligence released Ripple-65B, a 65 billion parameter dense transformer under an Apache 2.0 license as a preview checkpoint, and its published benchmarks show it leading terminal-based agentic coding and a harder software engineering benchmark while trailing Meta and Reflection-branded models on math, safety, and automation tests. The model activates all 65 billion parameters on every token, uses a thin-and-deep Gemma-style layout with 80 layers and sliding-window attention, and was pre-trained on 12 trillion tokens of mostly natural English text and code with synthetic data held back for later stages. Recursive Intelligence said the preview received light post-training concentrated on agentic coding, so the numbers reflect an early checkpoint rather than a fully tuned model.", "body_md": "# Ripple-65B Benchmarks: A Dense Model That Punches Above Its Weight\n\nRipple-65B is a dense 65B model from Recursive Intelligence that leads on agentic coding benchmarks but trails on math and safety tests.\n\n## What is Ripple-65B?\n\nRipple-65B is a 65 billion parameter dense language model built by a small team at Recursive Intelligence, released under an Apache 2.0 license as a preview checkpoint. It is trained from scratch, aimed specifically at agentic coding and reasoning tasks, and distinguished mainly by one architectural choice: it is dense, not mixture of experts (MoE), which runs counter to how most frontier labs are building large models right now.\n\n## TL;DR\n\n- **Ripple-65B is a dense transformer** , meaning all 65 billion parameters activate on every token, unlike MoE models that only wake a subset of experts per token.\n- On **agentic coding benchmarks** , the model reportedly leads in terminal-based tasks and a harder software engineering benchmark, outperforming models several times its size.\n- On a **deep software UI benchmark** , Ripple-65B lands near the top tier but sits behind Meta’s comparison model.\n- The model is noticeably weaker on **math, safety, and automation benchmarks** , where Meta and Reflection-branded models take the lead.\n- Architecturally, Ripple is **thin and deep** , with 80 layers and a narrow width, using a Gemma-style layout, extra attention heads, and sliding-window attention layers to control inference cost.\n- Pre-training ran on **12 trillion tokens** of mostly natural English text and code, with synthetic data deliberately held back for later training stages.\n- This is a **preview release with light post-training** , so the published benchmark numbers reflect an early checkpoint rather than a fully tuned final model.\n\n## How does Ripple-65B perform on agentic coding benchmarks?\n\n*Rapidly*build enterprise-grade software applications.\n\nThe headline result is coding performance. Across the benchmark charts shared by Recursive Intelligence, Ripple-65B’s score (shown as the dark blue bar) comes out on top in a terminal-based agentic benchmark and in a harder software engineering benchmark, beating models with considerably more parameters. That is the core claim behind the model: a comparatively small dense network holding its own, or winning outright, against much larger competitors on tasks that involve writing, running, and fixing code in an agent loop rather than answering a single static question.\n\nOn a separate “deep software UI” style benchmark, Ripple-65B sits in the top cluster of models but does not lead. Meta’s comparison model is clearly ahead on that particular test, which suggests Ripple’s strength is more concentrated in terminal and code-repair style agentic tasks than in broader UI or interaction benchmarks.\n\n## Where does Ripple-65B fall short?\n\nThe same benchmark set shows clear weak points. Math, safety, and automation are the categories where Ripple-65B trails, and where Meta and Reflection-branded models take the lead instead. This lines up with how the model was trained: Recursive Intelligence has said this preview received relatively light post-training, with the training effort concentrated on agentic coding rather than broad general capability. Math reasoning and safety alignment typically benefit heavily from extensive post-training and RLHF-style tuning, so a model that skipped deep investment there scoring lower is consistent with the stated training approach rather than a surprise.\n\nThe practical read: Ripple-65B looks like a specialist, not a generalist. It is a coding and reasoning model first, with other capabilities trailing behind because the team’s limited post-training budget went toward the agentic coding pipeline specifically.\n\n## Why build a dense model instead of MoE?\n\nNearly every major lab shipping large models in 2024 and 2025 has leaned toward [mixture of experts architectures](https://goremy.ai/blog/articles/best-ai-models-for-internal-tools-dense-vs-moe-vs-reasoning), where only a fraction of total parameters activate for any given token. This keeps inference cheap relative to total parameter count and lets labs scale total capacity without scaling compute cost as fast.\n\nRipple goes the other direction. All 65 billion parameters activate on every token. The team’s stated reasoning is practical: dense training is simpler to get right with a small team. MoE architectures introduce routing complexity, load-balancing problems, and training instability that require significant infrastructure and tuning experience to manage well. A dense model sidesteps that complexity entirely, trading some inference efficiency for a more predictable training process that a small team can execute without the overhead of building MoE-specific tooling.\n\n## What does the “thin and deep” architecture actually mean?\n\nRipple-65B uses 80 layers with a narrow width per layer, rather than the wider, shallower designs common in many dense models of similar size. A few specific choices define the architecture:\n\nThe model adopts a Gemma-style layout for training stability at this depth. It uses more attention heads than typical to help with reasoning quality. It also mixes in more sliding-window attention layers with a smaller window size, which caps the computational cost that would otherwise come from running full attention at 80 layers deep. On top of that base, the architecture adds gated attention and value residual connections, both aimed at improving training dynamics without inflating parameter count further.\n\nThe combination amounts to a bet that depth, not width, is the more efficient way to get reasoning capability out of a fixed parameter budget, at least for a model this size trained by a small team.\n\n## How was Ripple-65B trained?\n\n*demands*of enterprises.\n\nRun it on our own servers.\n\n**✓**\n\nSOC 2 and HIPAA from day one.\n\n**✓**\n\nLog every action.\n\n**✓**\n\nOnly the right people get in.\n\n**✓**\n\nPre-training ran on 12 trillion tokens, drawn mostly from natural English text and code rather than heavily synthetic data. Recursive Intelligence has said they kept this initial corpus clean and deliberately held synthetic data back, reserving it for later training stages so they could iterate on synthetic data quality without blocking or slowing down the pre-training run itself.\n\nAfter pre-training, the model went through a multi-stage post-training chain: a mid-training stage, a long-context stage, supervised fine-tuning, a second long-context fine-tuning pass, and finally reinforcement learning across mixed domains. Recursive Intelligence describes this current release as having gone through relatively light post-training, with the focus placed specifically on agentic coding capability. That framing matters for interpreting the benchmark results: the coding strength is a direct product of where the limited post-training effort was spent, and the weaker math and safety scores reflect where it wasn’t.\n\n## Is Ripple-65B worth running locally?\n\nFor teams interested in self-hosting, Ripple-65B’s size puts it in reach of serious but not exotic hardware. It is designed to fit on devices like NVIDIA’s DGX Spark or similar systems, and quantized versions can run on a single high-VRAM GPU setup (one independent test ran a quantized version using around 75GB of VRAM). That’s accessible for a well-equipped workstation or small server, though it’s well beyond consumer GPU territory.\n\nSetup is not necessarily plug-and-play. Independent testing found that the provided inference code was built around a specific GPU target and needed adjustment to run on other hardware, and getting a full local deployment working (download, configuration, inference server, agent harness) took several hours even for someone experienced with local model hosting. For teams without in-house ML infrastructure experience, that setup cost is a real factor to weigh against the benchmark gains.\n\nThe Apache 2.0 license removes the usual commercial licensing friction, which is a meaningful point in its favor for teams wanting to fine-tune or deploy commercially without negotiating terms.\n\n## Frequently Asked Questions\n\n### What company made Ripple-65B?\n\nRipple-65B was built by Recursive Intelligence, described as a small team that trained the model from scratch rather than fine-tuning an existing base model.\n\n### Is Ripple-65B a mixture of experts model?\n\nNo. Ripple-65B is a dense transformer, meaning all 65 billion parameters are active for every token processed, unlike MoE architectures that route each token through only a subset of the model.\n\n### What license is Ripple-65B released under?\n\nIt’s released under Apache 2.0, an open license that allows commercial use, modification, and redistribution without the restrictions found in more limited research-only licenses.\n\n### What is Ripple-65B’s biggest weakness according to its benchmarks?\n\nMath, safety, and automation benchmarks are where the model scores lowest relative to competitors, trailing models from Meta and Reflection-branded releases in those specific categories.\n\n### How much hardware does it take to run Ripple-65B locally?\n\nIt’s sized to run on systems like NVIDIA’s DGX Spark, and quantized versions have been run on a single high-VRAM GPU setup using roughly 75GB of VRAM, though inference code may need adjustment depending on the specific GPU used.", "url": "https://wpnews.pro/news/ripple-65b-benchmarks-a-dense-model-that-punches-above-its-weight", "canonical_source": "https://www.mindstudio.ai/blog/ripple-65b-benchmarks-coding/", "published_at": "2026-11-05 00:00:00+00:00", "updated_at": "2026-11-05 10:54:57.439954+00:00", "lang": "en", "topics": ["large-language-models", "artificial-intelligence", "ai-research", "ai-products"], "entities": ["Recursive Intelligence", "Ripple-65B", "Meta", "Reflection", "Gemma"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/ripple-65b-benchmarks-a-dense-model-that-punches-above-its-weight", "markdown": "https://wpnews.pro/news/ripple-65b-benchmarks-a-dense-model-that-punches-above-its-weight.md", "text": "https://wpnews.pro/news/ripple-65b-benchmarks-a-dense-model-that-punches-above-its-weight.txt", "jsonld": "https://wpnews.pro/news/ripple-65b-benchmarks-a-dense-model-that-punches-above-its-weight.jsonld"}}