Ricursive Intelligence released Ripple-65B Preview on November 4 — a 65-billion-parameter dense model that scores 80.52 on TerminalBench 2.1, beating Reflection Beam’s 501B and Muse Spark Open’s 816B on the benchmark that matters most for coding agents. It is Apache 2.0, ungated on Hugging Face, and fits on a single NVIDIA DGX Spark at roughly 57 GB (NVFP4). The company was founded by the researchers who built AlphaChip for Google. This is worth your attention.
Dense, Not MoE — and That Is the Point
Every major coding model released in the past six months has been mixture-of-experts: DeepSeek V4 Flash, Qwen3-Coder, Mistral Large 4, Muse Spark Open. Ricursive chose dense, and not by accident. When every parameter fires on every token, you get lower-latency generation and dead-simple deployment — no custom MoE sharding kernels, no expert routing bugs, no specialised infrastructure. A small team can train it, and a single well-specced workstation can run it.
The trade-off is real. Ripple-65B trails on knowledge breadth and multilingual tasks. It loses on MCP Atlas (71.67 vs. Muse Spark Open’s 89.2) and on scientific code generation. It was not built to be a generalist. It was built to write, run, and fix code inside an agent loop — and on that specific task, it currently leads among open-weight models at its size.
Benchmark Numbers (Read the Footnotes)
On TerminalBench 2.1, Ripple-65B scores 80.52. For context:
| Model | TerminalBench 2.1 | Parameters | Architecture |
|---|---|---|---|
| DeepSeek V4 Flash | 88.6 | 284B | MoE |
| Qwen 3.8 Flash Next | 84.3 | undisclosed | MoE |
| Ripple-65B Preview | 80.52 | 65B | Dense |
| Reflection Beam | 80.1 | 501B | MoE |
| Muse Spark Open | 76.4 | 816B | MoE |
Ripple also posts 79.1 on SWE-bench Pro v2.0 Hard — the best open-weight result at launch. These numbers are vendor-run. Ricursive ran the evaluations, not an independent lab. That does not make them wrong, but it does mean you should run the model on your own agentic coding tasks before making pipeline decisions based on these scores.
How to Get It
Ripple-65B is fully open under Apache 2.0. No form to fill in, no approval queue, no usage restrictions beyond the license itself. The model card and weights live at ricursiveAI/Ripple-65B-Preview on Hugging Face.
If you want to run it locally, the NVFP4 checkpoint is about 57 GB and runs on a single DGX Spark with 128K context. The BF16 version is roughly 130 GB and needs two B200 GPUs for full 1M-token context (250 tokens per second at batch size 8). Setup uses SGLang 0.5.19 with tensor parallelism on Linux with CUDA 13.
If you want to test it without the hardware overhead, it is available through Together AI, OpenRouter, GMI Cloud, and Fireworks. OpenRouter pricing is $0.20 per million input tokens and $0.60 per million output — competitive for a model at this performance tier. One caveat: the hosted API caps context at 16,384 tokens, not the architectural 1M. That is a KV cache memory constraint at the providers, not a model limitation, but it matters if your use case depends on very long context.
from transformers import AutoTokenizer, AutoModelForCausalLM
tokenizer = AutoTokenizer.from_pretrained("ricursiveAI/Ripple-65B-Preview")
model = AutoModelForCausalLM.from_pretrained(
"ricursiveAI/Ripple-65B-Preview",
torch_dtype="bfloat16",
device_map="auto"
)Who Built This
Ricursive Intelligence was founded in December 2025 by Anna Goldie and Azalia Mirhoseini, two Google Brain researchers who built AlphaChip — a deep reinforcement learning system that designs chip layouts. Their 2020 Nature paper demonstrated that RL could produce chip floorplans competitive with human engineers in a fraction of the time. Google used it on several TPU generations. Both were named to Time’s 100 Most Influential People in AI for 2026.
The company raised a $300 million Series A — NVIDIA Ventures participated — at a $4 billion valuation. The name is deliberate: recursive self-improvement, the idea that better AI produces better chips, which power better AI. Ripple-65B cost roughly $32 million in compute to train, which Ricursive describes as about a tenth of what the previous open-weight coding leader spent.
Where It Fits
Ripple-65B is not the best coding model available. DeepSeek V4 Flash leads TerminalBench at 88.6, and Qwen 3.8 Flash Next is ahead at 84.3. What Ripple offers is a compact, Apache 2.0 dense model that beats much larger MoE models on the benchmarks that matter most for coding agents, runs on accessible hardware, and comes from a credible team with a clear architectural thesis.
If you are building agentic coding pipelines and want an open-weight model you can deploy commercially without restrictions, Ripple-65B deserves a spot in your evaluation queue. Run it on your own tasks. Treat the vendor benchmarks as a starting point, not a verdict.
