Ripple-65B: US startup trains an open coding model for just $32 million

What's going on

US company Ricursive Intelligence released Ripple-65B Preview on November 4, 2026. The language model has 65 billion parameters, and the company says it trained it entirely in-house. In other words, it isn't built on top of a model from Meta, Alibaba or anyone else. The weights are freely available on Hugging Face at ricursiveAI/Ripple-65B-Preview under the Apache 2.0 license, one of the most permissive licenses around: you can use the model commercially, modify it and redistribute it. Ripple is also available via API from Together AI, GMI Cloud and OpenRouter. According to startup portal ecosistemastartup.com, Fireworks added it the same day.

Ricursive doesn't see itself as a typical model lab. It calls itself a "full-stack lab" that co-optimizes hardware, software and models, with the goal of building self-improving systems. Ricursive says Ripple is the general-purpose version of an unreleased in-house model that powers the company's chip design platform.

The claim: small beats big at agentic coding

Ricursive is pitching Ripple as the new leader among open US models for agentic coding. That means the AI works on its own in a terminal: it runs commands, edits files and handles multi-step tasks. The main yardstick here is TerminalBench 2.1. Here's the table the company published:

The Model Card lists more results: 79.1 on SWE-bench Pro v2.0 Hard, 76.5 on SWE-bench Verified and 44.3 on DeepSWE 1.1. Every one of these numbers comes from Ricursive. According to OrcaRouter, the Model Card quietly notes that the benchmark scores were "hard to reproduce." They were measured at maximum effort and full BF16 precision. Nobody has independently verified them yet.

How the model is built

Most labs use Mixture-of-Experts (MoE). In an MoE model, only a small part of the network does the work for each token, a bit like a big hospital where a patient only sees the relevant specialists. Ricursive deliberately went the other way. Ripple is a dense model, so all 65 billion parameters are active for every token. The company says this makes training and serving simpler. It claims it sustained 40 percent model FLOPs utilization (MFU) with straightforward sharding across chips, no custom GPU kernels and no quantization.

The architecture is "thin and deep": 80 layers with a width of 5,120. Ricursive describes the design as a move away from the Allen Institute for AI's open Olmo 3 and toward a layout similar to Google's Gemma 3N. It includes:

• more attention heads than usual. Attention is the mechanism the model uses to decide which earlier parts of the text matter right now.
• a higher share of sliding-window attention. These layers only look at a limited stretch of text, like a reader who only remembers the last few pages. That makes them much cheaper than global attention, which looks at the whole text.
• "gated attention," global NoPE (layers without positional encoding), value residuals and clamping of extreme values in the feed-forward layers.

Pretraining ran for 877,500 steps on 8,000-token sequences, 12 trillion tokens in total. The data is almost entirely English text and code. Then came mid-training, a long-context phase and a light round of reinforcement learning (RL), where the model learns by getting rewarded for solving tasks. That phase focused on agentic coding. The model is text-only and doesn't understand images.

Cheap and efficient to train, according to the company

The second big headline is the price. Ricursive says the entire project cost about $32 million. That figure supposedly covers early experiments, data mixing, evaluation, debugging, synthetic training data and the final training runs. The company says that's roughly a tenth of what the previous leader cost. For comparison, it points to Sam Altman, who said in 2023 that GPT-4 cost more than $100 million. Of rival Reflection, Ricursive writes that it reportedly committed $150 million a month to compute. TNW also reported compute deals worth $6.3 billion with SpaceX and $1 billion with Nebius.

For RL, Ripple reportedly needed only 1.4 million rollouts, practice runs where the model attempts a task and gets scored. The company says that's about 70 times fewer than Reflection Beam. Instead of buying training data, Ricursive generated and filtered synthetic data with its own pipelines. It started from Olmo 3's experiment series (the "ladder"). Over six iterations (v1.0 to v1.6), Ricursive says it improved compute efficiency for code training by up to 16x, measured on a smaller 4-billion-parameter model. Ecosistemastartup.com sees this as the part of the story others can actually copy: run cheap tests at small scale first, and only then train big.

Pricing, context and contradictions

According to OrcaRouter, third-party providers charge around $0.20 per million input tokens and $0.60 per million output tokens for Ripple. For comparison, Kimi K3.1 Pro, released the same day, costs $2 and $10. Ricursive hasn't published its own price list, though. And the first hosted endpoint offers only 16,384 tokens of context instead of the advertised 1 million. OrcaRouter also notes that at least one tracker wrongly lists the model as proprietary.

OrcaRouter also warns about a popular apples-to-oranges comparison. On forums, Ripple's 80.5 on TerminalBench 2.1 often gets put next to Kimi's 65.7 on TerminalBench 4.0. But version 4.0 is a later, harder test, so the two numbers can't be compared. Even for the same model on the same test, results can swing by several points depending on the test harness.

What you need to try it

At full BF16 precision, the model weighs about 130 GB. That won't fit on any single consumer GPU, and despite what some reports suggest, it won't easily fit on a MacBook either. Ricursive names NVIDIA DGX Spark as the target device, a compact AI machine with lots of unified memory. On two B200 data center GPUs, the company says the model hits 250 tokens per second with 8 concurrent requests. The repository also includes an NVFP4 variant, a 4-bit format for Nvidia's newer Blackwell generation that makes running it on smaller machines possible in the first place.

As a rough rule of thumb, a 4-bit version needs about 40 GB of VRAM (video memory). That means two 24 GB cards, a workstation card, or a Mac with 64 GB or more of unified memory. The sources don't mention ready-made GGUF files for llama.cpp or Ollama yet. If you don't have the hardware, you can try the model cheaply on OpenRouter or Together AI. The longer the context, the more memory you'll need on top.

The bigger picture

Ripple is interesting for two reasons. First, according to CNET, the strongest open models currently come from China: DeepSeek, Kimi K3 and Qwen3.8 top the Artificial Analysis leaderboards. A fully home-grown, relatively small US model under a genuine open-source license is a statement. Second, if the numbers hold up, Ricursive shows that strong agentic performance doesn't necessarily require billion-dollar budgets and trillion-parameter models.

Still, be careful. Almost all the numbers come from the company. The Model Card admits reproducibility problems, and this is a "preview" release. The model is focused on English and code, so expect weaker results in other languages. But if you want to run a local coding agent on your own hardware or fine-tune a model, a dense 65B model under Apache 2.0 is a far more accessible starting point than a trillion-parameter MoE.