Ricursive releases Ripple-65B Preview, a 65B open-weight model for agentic coding

Ricursive releases Ripple-65B Preview, a 65B open-weight model for agentic coding
📅 November 6, 2026 • The Decoder

What was released
Ricursive Intelligence has published Ripple-65B Preview, a dense Transformer built for agentic coding work. The preview ships under the Apache 2.0 licence, so the weights and the code can be used, modified and redistributed without a separate commercial agreement. The model is listed on Hugging Face and is served through Together AI, OpenRouter, Fireworks and GMI Cloud. Together AI lists 0.20 dollars per 1 million input tokens and 0.60 dollars per 1 million output tokens, so the release is reachable at both the entry and the mid tier of the market.

How the model is built
Ripple-65B is a dense architecture rather than a mixture of experts, which means every token is processed through the full parameter set. Training data covered 12 trillion tokens and the context window reaches 1 million tokens. The long window is the practical part for agentic coding, where a task usually requires a repository, several logs and a ticket thread to stay readable at the same time while the assistant keeps editing files.

Reported benchmark results
The headline number is 80.52 on TerminalBench 2.1, a benchmark for terminal and command line tasks. Per Ricursive, that figure sits above Reflection Beam 501B at 80.1, Meta Muse Spark Open 816B at 76.4, Inkling 975B at 63.8 and Nemotron 3 Ultra 550B at 56.4. The company describes Ripple as the top result for a United States open-weight model on that benchmark. The Decoder's comparison table lists 79.1 percent on SWE-Bench Pro v2.0 Hard and 71.67 percent on MCP Atlas for Ripple-65B Preview. These are vendor reported figures produced by the team that built the model.

Cost and training efficiency
Ricursive reports total research, development and training cost of about 32 million dollars. Reaching the TerminalBench score took 1.4 million reinforcement learning rollouts, which per Ricursive is roughly 70 times fewer than the run behind Reflection Beam, the previous state of the art result. Fewer rollouts mean less generated training traffic and a shorter reinforcement learning phase, which is usually the most expensive part of the pipeline for a model of this size.

What it means in practice
For teams choosing an open-weight coding model, Ripple-65B Preview offers a combination that is hard to find together: a permissive licence, a long context window, published benchmark numbers and access through several inference providers. The caveats are the usual ones for a preview. Terminal agent benchmarks are sensitive to the scoring harness, provider pricing can change, and the model targets agentic coding rather than general conversation. The natural next step is to run the repository tasks a team already has and compare the results with the stack in use today.

IMPACT Ripple-65B Preview gives open-weight coding teams a permissive licence, a 1 million token context and published agent benchmarks, so the next question is how the preview behaves on real repository work rather than on a leaderboard.

Sources: https://the-decoder.com/ripple-65b-preview-delivers-strong-agentic-coding-results-as-a-fully-apache-2-0-open-us-model/ — https://ricursive.com/ripple — https://huggingface.co/ricursiveAI/Ripple-65B-Preview — https://www.together.ai/models/ripple-65b-preview