Ricursive Intelligence shipped Ripple-65B Preview on November 4 — a 65-billion-parameter dense open-weight model that now tops every US open-weight model on two major agentic coding benchmarks, beating MoE rivals with up to 975 billion parameters. The model is Apache 2.0, runs locally on a DGX Spark, and is live on OpenRouter and Together AI right now.
The Numbers, Honestly
Ripple-65B scored 80.52 on TerminalBench 2.1, an execution-based benchmark where agents navigate real Bash environments, write code, manage dependencies, and handle system administration tasks. That edges Reflection Beam (80.1 at 501B parameters) and beats Meta Muse Spark Open (76.4 at 816B) and NVIDIA Nemotron 3 Ultra (56.4 at 550B) by a wider margin. On SWE-bench Pro Hard, Ripple scores 79.1 against Beam’s 76.5.
Worth saying plainly: the TerminalBench lead over Beam is 0.42 points, and the comparison is scoped to US-origin open-weight models — Qwen models are excluded. If you need the absolute best open-weight coder regardless of origin, that benchmark story is different. If you are in an air-gapped or regulated environment that requires a US-provenance model, this is currently your best option by a clear margin.
How to Use Ripple-65B Today
The fastest path is the API. Ripple is available on OpenRouter via GMI Cloud at $0.20 per million input tokens and $0.60 per million output tokens, with $0.05 per million for cache reads. Together AI also carries the model. The API supports native tool calling via tools and tool_choice parameters, and response_format for JSON output. Note that the API context window is 16,384 tokens — not the 1 million figure in the marketing materials.
import openai
client = openai.OpenAI(
base_url="https://openrouter.ai/api/v1",
api_key="<your-openrouter-key>",
)
response = client.chat.completions.create(
model="ricursiveai/ripple-65b-preview",
messages=[{"role": "user", "content": "Fix the memory leak in this Go code: ..."}],
tools=[...], # native tool calling supported
)For local deployment, GGUF files are available at AtomicChat/Ripple-65B-Preview-GGUF on Hugging Face. There is a catch: Ripple’s architecture is not yet in upstream llama.cpp. PR #31692 is open but unmerged, so you will need a patched build for now. If you have a DGX Spark, Ricursive provides an official NVFP4 checkpoint with a Colab notebook — that is the cleanest local path currently.
Why a 65B Dense Model Beats 975B MoE Rivals
MoE models activate only a subset of parameters per token, which makes them efficient at scale but adds routing complexity. For agentic task loops — where an agent calls tools, reads results, and loops repeatedly — inference latency per step compounds fast. A dense model with no routing overhead runs each step faster, even if its throughput at high batch sizes is lower than a comparably-capable MoE.
Ricursive leaned into this. Ripple was co-designed with NVIDIA DGX Spark in mind, using a thin-and-deep architecture (80 layers, narrow 5,120-dimensional embeddings) that fits and runs efficiently on that hardware. The $32M training cost — roughly one-tenth the compute cost of the prior US open-weight SOTA — reflects the same efficiency-first mindset applied to training.
Benchmark Comparison
| Model | TB 2.1 | SWE-bench Pro Hard | Parameters |
|---|---|---|---|
| Ripple-65B | 80.52 | 79.1 | 65B dense |
| Reflection Beam | 80.1 | 76.5 | 501B MoE |
| Meta Muse Spark Open | 76.4 | 72.4 | 816B MoE |
| Thinking Machines Inkling | 63.8 | 39.6 | 975B MoE |
| NVIDIA Nemotron 3 Ultra | 56.4 | 30.7 | 550B MoE |
Who This Is For
Ripple makes the most sense for three groups. First, teams in regulated, air-gapped, or government environments where US-origin provenance matters for procurement or compliance. Second, developers building agentic coding pipelines who want native tool calling and do not want to manage a 500B-parameter system. Third, researchers and teams with DGX Spark hardware who want a frontier-capable coder optimized for that specific setup.
If none of those apply, Qwen3.6-27B competes on a number of coding benchmarks at roughly one-quarter the parameter count, and Qwen is not restricted to US provenance. Use Ripple when provenance or the DGX Spark integration specifically matters to you.
The Company
Ricursive was founded by Anna Goldie and Azalia Mirhoseini, the researchers behind AlphaChip at Google Brain. Their original pitch was AI for chip design — compressing design cycles from years to weeks. Ripple-65B is their first language model, and it carries that same theme: a chip-design shop that co-designs its AI models with the hardware they intend to run on. The model is available now, it is Apache 2.0, and the weights are on Hugging Face.
