# Ripple-65B Ripple-65B Preview is a 65B [[Dense AI Models|dense]] open-weight model from Ricursive Intelligence, published on Hugging Face around 4 November 2026 under [[Apache 2.0 License|Apache 2.0]]. The headline: "a new U.S. open-weight SOTA on agentic coding", with 80.5 on Terminal-Bench 2.1. Ricursive isn't a model lab first. It builds an AI chip design platform, and describes itself as "a full-stack artificial intelligence lab building self-improving systems by co-optimizing hardware, software, and model". Ripple is "a general-purpose variant of the proprietary model that powers our AI chip design platform". ## What shipped - **65B parameters, all active.** A dense Transformer, 80 layers, d_model 5120. Ricursive chose dense over [[AI Mixture of Experts (MoE)|MoE]] on purpose - **Trained from scratch** on 12T tokens, with up to 1M tokens of context - **Text only**, mostly English and code - **BF16 weights**, plus a separate NVFP4 build - **Apache 2.0, no extra conditions.** No revenue threshold, no attribution rule (compare with [[Kimi K3.1 Pro]]) - **Hosted** on Together AI, GMI Cloud, [[OpenRouter]] and [[Fireworks AI]] - **Runs on one box.** Ricursive says it fits on an NVIDIA DGX Spark, and reports about 250 tokens/s on 2x B200 at batch size 8, before quantization or custom kernels ## The cost claim $32M of total R&D, everything included: the pretraining ladder, data mixing, evals, synthetic data, post-training and the final run. Ricursive calls that about a tenth of the previous open-weight SOTA (its own estimate). A few details explain how: - No post-training data was bought. Synthetic pipelines and quality filtering instead - About 40% MFU (model FLOPs utilization: how much of the hardware's peak compute training actually used) with "simple sharding, no custom kernels, and no quantization" - 1.4M RL rollouts, which they say is about 70x fewer than Reflection Beam - Architecture tweaks: gated attention, global NoPE, value residuals, and more sliding-window layers with a smaller window ## Benchmarks The comparison Ricursive chose, Terminal-Bench 2.1, US open-weight models only: | Model | Parameters | Terminal-Bench 2.1 | |---|---|---| | **Ripple-65B Preview** | 65B dense | **80.5** | | Reflection Beam | 501B | 80.1 | | Meta Muse Spark Open | 816B | 76.4 | | Thinking Machines Inkling | 975B | 63.8 | | [[NVIDIA Nemotron]] 3 Ultra | 550B | 56.4 | The model card adds SWE-bench Verified 76.5, SWE-bench Pro v2.0 Hard 79.1, AIME 2025 98.3, GPQA Diamond 85.2 and IFBench 79.2. Most of those come from Ricursive's own setup, and the card itself warns that agentic scores are hard to replicate and depend on the harness. ## What Hacker News pointed out The thread (339 points) mostly picked apart the framing: - **The chart picks its rivals.** "U.S. open-weight" leaves out every Chinese model. [[Qwen3.8-27B]], also dense and 2.4x smaller, scores 73.0 on the same benchmark and isn't on the chart. Ripple still beats it by 7.5 points, so the claim holds, but it's much narrower than the headline. As wmf put it: "a pretty bad look to not even mention those other open weight models" - **An older benchmark.** Terminal-Bench 4.0 is the current version (the one [[Kimi K3.1 Pro]] is measured on) - **Total parameters flatter it.** namjh noted that comparing a dense 65B with MoE models by total size is misleading, since the rivals activate far fewer parameters per token - **Dense is slow on consumer hardware.** Every token uses all 65B. The counter-argument: [[AI Speculative Decoding]] (multi-token prediction drafts, DFlash) makes dense models practical again, and Mac Studio owners were happy to have a strong model that fits in their memory. stewartpark still wanted "150tk/s instead of 50tk/s!" - **Tool calls failed.** pcwelder tested a hosted version and saw 4 of 8 tool calls fail: "the tool call failure rate is not acceptable." One user, one provider, a preview model. But for an agentic coding model, that's the number to check first - **No chip design benchmark**, from a chip design company ## Who it's for Several commenters said their companies or government clients ban Chinese models outright. mcnewcp: "they still refuse to let me deploy anything from China." stewartpark's guess at the market: "US Gov't agencies, chip design industry". dragonwriter pushed back: with open weights, origin gives you no reason to trust one model more than another, since bias and backdoors could sit in either. aftbit replied that switching costs and vendor relationships still matter when you need the next version. ## My take I find the cost more interesting than the SOTA claim. $32M for a competitive model trained from scratch is within reach of a large company or a government agency. If that number holds up, "train our own" becomes a real option, which is [[Sovereign AI]] at the organizational level. The origin argument is weaker than procurement rules make it look. You can run open weights offline and inspect them, so the risk with a Chinese model is hidden behavior, not data leaving the building. But procurement rules don't do nuance, and a US-made, Apache 2.0 model at this level is exactly what those buyers didn't have. For local agentic coding on a single 24 GB GPU, I'd still pick Qwen3.8-27B: Ripple needs roughly 35 GB of weights even at 4-bit, and the tool-call reports need confirming first. On a Mac with enough unified memory, it's worth a try. ## References - Ricursive Intelligence, "Ripple-65B Preview": https://www.ricursive.com/ripple - Model card: https://huggingface.co/ricursiveAI/Ripple-65B-Preview - Hacker News, "Ripple-65B Preview: U.S. Open-Weight SOTA on Agentic Coding" (339 points): https://news.ycombinator.com/item?id=50369925 ## Related - [[Dense AI Models]] - [[AI Mixture of Experts (MoE)]] - [[AI Speculative Decoding]] - [[Qwen3.8-27B]] - [[Kimi K3.1 Pro]] - [[NVIDIA Nemotron]] - [[AI Open Weight Models]] - [[Apache 2.0 License]] - [[Sovereign AI]] - [[Benchmaxxing]] - [[OpenRouter]] - [[Fireworks AI]]