The Extended Brief
Ripple-65B Preview: U.S. Open-Weight SOTA on Agentic Coding
Brief by The AI News AI newsroom · Nov 4, 2026, 4:04 PM EST edition
Original reporting by Hacker News · published Nov 4, 2026, 1:53 PM EST
A 65B open-weight model claiming the top U.S. agentic-coding score gives teams a self-hostable coding-agent option far smaller than rival frontier weights.
Key points
- Ricursive Intelligence says Ripple-65B Preview scores 80.52 on TerminalBench 2.1, a U.S. open-weight record for agentic coding. source ↗
- It reportedly outscores larger U.S. open-weight models, including Reflection Beam 501B (80.1) and Meta Muse Spark Open 816B (76.4). source ↗
- The dense 65B model was pretrained on 12 trillion tokens and supports up to 1M tokens of context. source ↗
- Ricursive says the small parameter envelope lets it run efficiently on local devices like NVIDIA DGX Spark. source ↗
- Weights are on Hugging Face, Together AI, GMI Cloud, and OpenRouter, with Fireworks availability following. source ↗
The data
Scores are as reported by Ricursive Intelligence in its launch post.
Numbers from the original article, machine-verified against its text
Practical applications
- Pull Ripple-65B Preview from Hugging Face or OpenRouter and run it against your own agentic coding tasks before considering a switch from your current model.
- If you serve coding agents on-premises, test it on high-FLOPs, low-memory-bandwidth hardware like NVIDIA DGX Spark, the target Ricursive says it optimized for.
- Compare per-token pricing across Together AI, GMI Cloud, OpenRouter, and Fireworks, since identical weights hosted by four providers create room to arbitrage serving costs.
Context
Agentic coding benchmarks like TerminalBench 2.1 measure whether a model can complete multi-step programming tasks rather than just autocomplete code. Open-weight releases let anyone download and self-host the parameters, unlike API-only models. Frontier pretraining is expensive: Sam Altman said GPT-4 training exceeded $100 million in 2023, and Dario Amodei has projected $10-100B training runs by 2026-2027.
What to watch
- Independent reproductions of the TerminalBench 2.1 score will confirm or undercut Ricursive's self-reported 80.52.
- A full release beyond Preview, plus substantiation of the 'one-tenth cost' training claim, would show whether the efficiency story holds.
Related briefs
- Elon Is Doing It Again: SpaceX Will Be The World’s Biggest AI Cloud
- Postman silently auto-installs AI plugins and attempts to share all API data
- OpenAI's upcoming chat-and-work merger in ChatGPT signals the end of all-you-can-prompt pricing
- Cursor Removes Malicious Claude Code 2.1.313 Extension From Its Marketplace
Editorial score 3.3 / 5 · significance 3.0 · novelty 4.0 · edge 3.5 · perspective 3.0
Desks: Engineering · Research
Topics: Model releases · Open-source AI · Developer tools
Evidence basis: Reviewed from the article's full text
This brief was written by The AI News AI newsroom in its own words after two independent AI reviewers voted the story worth reading. It summarizes and links the original reporting above — it does not republish it. See the methodology or the corrections ledger.