Ricursive Intelligence has released Ripple-65B Preview, a 65-billion-parameter open-weight model designed for agentic coding and AI chip design. The company says it scored 80.52 on TerminalBench 2.1, which it describes as a US open-weight state of the art.
Watch Desk analysis
What happened
Ricursive says Ripple-65B was trained from scratch for $32 million, using a dense Transformer architecture and custom reinforcement-learning methods. The company says the model is available through Hugging Face and Together AI.
The release is aimed at two demanding jobs: coding agents and chip design. Ricursive also claims Ripple outperforms significantly larger models on agentic coding tasks. The supplied announcement does not name those models or give a comparison table, so that claim is best treated as the company’s account, not a settled leaderboard verdict.
Why it matters
A 65-billion-parameter open-weight model aimed at chip design brings AI model development and the design of AI hardware into the same story. If the model performs as Ricursive says, it could give developers another option for coding agents and specialised engineering work, without relying solely on closed services.
The benchmark score offers a concrete starting point for comparison, but one result cannot tell us how well Ripple handles real engineering workflows, or how its performance changes with the hardware and software available to users. The useful next step is independent testing, not treating “state of the art” as a magic spell.
Our read
This is a substantial model release with a distinctive focus, rather than another general-purpose launch wearing a lab coat. The open-weight availability makes it worth watching, while the chip-design ambition is the more unusual part. For now, Ricursive’s benchmark and performance claims need testing against other models and practical tasks.
What to watch
- Independent TerminalBench 2.1 results and comparisons with other open-weight models.
- Whether Ripple shows useful gains on real chip-design tasks, beyond its stated optimisation target.
- What hardware is needed to run the model and how its open weights are licensed.
Discussion spark: For a model aimed at chip design, what would persuade you it is genuinely useful: independent benchmark gains, evidence from real design work, or access to the model weights?
Sources and evidence
- Recursive Self-Improvement via AI for Chip Design & Chip Design for AI – Ricursive Intelligence (4 November 2026, 18:53 UTC)
- Together Compiler: Continuous Compilation for AI inference – Together AI (4 November 2026, 18:53 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.