Notes on film, new tools and the decisions behind the work.

Tools / Analysis

Ripple-65B Preview brings an open-weight coding model to local hardware

Ricursive released a dense 65-billion-parameter model under Apache 2.0. Its coding results deserve a close look at the evaluation setup; the smaller checkpoint has a separate experimental status.

A hand at a desktop computer with a locally running coding model shown on the screen. Conceptual editorial illustration, not a benchmark, product photograph or tested hardware configuration.
DU / SIGNALS · S/099Editorial illustration · AI-generated

Ricursive Intelligence released Ripple-65B Preview on 4 November: a dense language model trained from scratch on 12 trillion tokens. Public weights are available under Apache 2.0, alongside hosted access. The company presents it as a general-purpose relative of the model behind its chip-design platform.

Ricursive names NVIDIA DGX Spark as a local target. That makes the release relevant to teams that want control over where their coding model runs, although fitting a checkpoint and sustaining a useful working context are separate deployment questions.

The local package differs from the benchmark model

The main BF16 repository lists about 130 GB of files. A separate NVFP4 checkpoint is marked experimental: it compresses the feed-forward layers while keeping attention in BF16. Quantization reduces the precision used to store parts of the model; the published BF16 benchmark scores should not be silently assigned to that smaller build.

Sources: Ricursive Intelligence / Hugging Face [3] · Ricursive Intelligence / Hugging Face [5] · Ricursive Intelligence / Hugging Face [2]

X / DEMO

Ricursive’s Ripple-65B Preview announcement

Loading post from X…

Credit / Ricursive Intelligence / @RicursiveAIOpen original 

Official release post. The linked model card supplies the benchmark conditions.

80.52 is a result with conditions

Ricursive reports 80.52 on Terminal-Bench 2.1, against a cited 80.1 for Reflection Beam. Its comparison uses BF16 at maximum effort, combining published vendor scores with in-house runs for missing results. The card warns that changing the agent harness and context handling changed scores. This is a company-reported comparison of selected US open-weight models, not a ranking of every model.

Sources: Ricursive Intelligence [1] · Ricursive Intelligence / Hugging Face [2]

A hosted trial before a local install

On 6 November, OpenRouter listed BF16 inference at $0.20 per million input tokens and $0.60 per million output tokens, with a 1,048,576-token context window. It supports tool calls and JSON output without JSON-schema enforcement. EU in-region routing was not available on the inspected listing.

For a team considering self-hosting, one useful first check is a representative repository task with a fixed test suite. Record the model revision, agent harness, context limit and failed tool calls. Those details make a hosted trial comparable with a later local run; the headline benchmark alone cannot do that.

Sources: OpenRouter [6]

SOURCES & METHOD

The record behind the article

Based on the official release, public model repositories, licence and provider listing. We have not run the model or reproduced the benchmarks.

  1. Introducing Ripple-65B Preview Ricursive Intelligence · 2026-11-04
  2. Ripple-65B Preview model card and evaluation conditions Ricursive Intelligence / Hugging Face
  3. Ripple-65B Preview public model files Ricursive Intelligence / Hugging Face
  4. Apache 2.0 licence in the released repository Ricursive Intelligence / Hugging Face
  5. Ripple-65B Preview NVFP4 — experimental checkpoint Ricursive Intelligence / Hugging Face
  6. Ripple-65B Preview API availability and pricing OpenRouter