Ricursive's Ripple-65B: Apache 2.0 open model that fits one DGX Spark

Ricursive Intelligence has released Ripple-65B Preview, an open-weight language model under the Apache 2.0 licence, according to the model card on Hugging Face. The company, which The Decoder describes as an AI chip design firm, reports that the model scores 80.52 on Terminal-Bench 2.1, a coding-agent benchmark, which it says is the top result among US open-weight models. The repository was created on 4 November 2026 and The Decoder reported the release on 6 November.

What it is

• A dense Transformer with 65 billion parameters, not a mixture-of-experts design. All 65B are active on every token, so it is smaller than the 500B-plus mixture-of-experts models it is compared with, but it does more work per token than they do.
• Pretrained on 12 trillion tokens of text, mostly English and code, with a context window of up to 1,048,576 tokens.
• Released as a BF16 instruction-tuned model, an NVFP4 4-bit checkpoint and a base model. Ricursive says post-training was relatively light and focused on agentic coding.
• Ricursive says total research and development cost was $32 million, as reported by The Decoder; the model card says a team that grew from one to five people trained it.

The numbers, and who ran them

Ricursive's table compares Ripple with four other US open-weight models. It reports Terminal-Bench 2.1 scores of 80.5 for Ripple, 80.1 for Reflection Beam 501B, 76.4 for Meta Muse Spark Open 816B, 63.8 for Thinking Machines Inkling 975B and 56.4 for NVIDIA Nemotron 3 Ultra 550B. Some cells use the labs' own model-card scores; others are Ricursive's in-house runs, marked in italics on the card.

The card is candid about the limits. It says benchmark scores are hard to replicate, and that Terminal-Bench 2.1 scores were lower for nearly every model when run with the official Terminus 2 harness at 128K sequence length. It is not leading everywhere: on its own table Ripple trails on SWE-bench Verified (76.5 against 81.6 for Reflection Beam), SciCode (39.4 against 58.5 for Muse Spark Open) and MCP Atlas (71.67 against 89.2 for Muse Spark Open). The Decoder adds that Inkling 2 Small, released around the same time, beats Ripple on the three benchmarks it tabulated.

Why a team running AI at work might care

The practical point is size. Ricursive says Ripple fits on hardware such as the NVIDIA DGX Spark and provides a 4-bit checkpoint and a Colab notebook to start. It reports 250 tokens per second on two B200 GPUs at batch size 8 before quantisation or custom kernels. For a team that wants a coding agent whose prompts and code never leave its own network, a model this size is a candidate to test.

Apache 2.0 also permits commercial use and modification, which many recent open-weight releases do not. That is the licence on the repository; read the LICENSE file there before you ship.

What the card leaves out

• This is a Preview. The card says post-training was light, and says nothing about safety testing beyond a few benchmark rows (FORTRESS, StrongREJECT) that Ricursive ran itself.
• No independent run of Terminal-Bench 2.1 for Ripple has been published that I could find.
• The card gives no cost per task, no latency under real agent loads, and no results on your language, repository size or tooling. Those are the numbers that decide a rollout.
• The repository ships custom code (modeling_ricursive.py), so it needs trust-remote-code style loading; have your platform team review it before you run it.