Ripple 65B Preview

An independent-style technical analysis reads Ricursive's launch post, Hugging Face model card, and OpenRouter listing, and separates documented facts from vendor-only measurements. It notes that Ripple-65B Preview is a 65B dense Transformer pretrained from scratch on 12 trillion tokens with a 1M-token context, published under Apache 2.0 and served through Together AI, GMI Cloud, and OpenRouter. The piece flags that the Terminal-Bench 2.1 score of 80.52 and the $32M build cost are vendor-run and unaudited, and that no independent benchmark of the model existed when it was reviewed on November 5, 2026, with the authors running none themselves. It also explains the unusual dense architecture choice and published cost figure.

A benchmark-oriented breakdown describes Ripple-65B Preview as a 65B-parameter dense open-weight Transformer pretrained from scratch on 12 trillion tokens with up to 1M context under Apache 2.0, released by Palo Alto startup Ricursive Intelligence, whose co-founders Anna Goldie and Azalia Mirhoseini previously worked on Google's AlphaChip. The release is a general-purpose variant of the model behind Ricursive's chip-design platform. The post stresses that the headline 80.52 Terminal-Bench 2.1 figure is self-reported, and reproduces the comparison table where Ripple leads Reflection Beam 501B (80.1), Muse Spark Open 816B (76.4), Inkling 975B (63.8), and Nemotron 3 Ultra 550B (56.4) on agentic coding while trailing several rivals on MMLU and AIME. It also notes the Preview label signals relatively light post-training, so it is an early checkpoint rather than a finished model.

The companion Hugging Face repository covers the NVFP4 variant of Ripple-65B Preview, marked experimental. The card describes it as a conservative quantization of the BF16 instruction-tuned checkpoint, applying NVFP4 to the MLP layers while keeping attention in BF16, and it lists a chat template and safetensors files for local deployment. That same page states the model is not deployed by any inference provider and asks for provider support, so the NVFP4 checkpoint currently relies on self-hosted or community serving. Its header also displays a 40B parameter figure that conflicts with the stated 65B dense design, so the metadata count should be treated as unverified.

The creator-published Hugging Face card for ricursiveAI/Ripple-65B-Preview documents the exact BF16 instruction-tuned checkpoint: a 65B dense Transformer pretrained on 12 trillion tokens with up to 1,048,576 tokens of context, released under Apache 2.0 with BF16, NVFP4, and base-model variants. The card also reports 80.52 on TerminalBench 2.1 against named US open-weight rivals. Training details are unusually specific: the team started from an Olmo 3 baseline and iteratively improved pretraining scaling laws, data, and post-training, claiming a 16x compute-efficiency gain on code over the Olmo 3 ladder. Ripple-65B was codesigned for NVIDIA hardware using a thin-and-deep aspect ratio (d_model 5120, 80 layers) with gated attention, FFN activation clipping, global NoPE, and value residuals, while holding 40% MFU with simple sharding and no custom kernels.

Ricursive Intelligence, a full-stack AI lab co-optimizing hardware, software, and model, released Ripple-65B Preview, a general-purpose variant of the proprietary model that powers its AI chip design platform. The model is a 65B dense Transformer pretrained on 12 trillion tokens with support for up to 1M context length, and it is available on Hugging Face, Together AI, GMI Cloud, OpenRouter, and Fireworks. On Ricursive's own measurements, Ripple-65B Preview scores 80.52 on Terminal-Bench 2.1, ahead of Reflection Beam 501B (80.1), Meta Muse Spark Open 816B (76.4), Thinking Machines Inkling 975B (63.8), and NVIDIA Nemotron 3 Ultra 550B (56.4). The company also reports 79.1 on SWE-bench Pro v2.0 Hard, 44.3 on DeepSWE 1.1, and 90.5 on HMMT Feb 2026, plus a claimed $32M total R&D cost and runs on devices such as NVIDIA DGX Spark.