Ivo Opens the Hood on Contract AI, Releasing Sage Model Built on DeepSeek

The San Francisco company built Sage in partnership with River AI, post-training it on DeepSeek V4 Flash — a foundation model that has gained traction among legal tech developers seeking high performance at lower inference cost. The release coincided with Ivo's Inscribe conference, signaling a deliberate push to build community around the product rather than treat it as a proprietary moat.

Legal AI has moved fast in 2026, but most enterprise tooling remains closed-source, leaving buyers to accept vendor benchmarks on faith. Ivo's open release breaks from that pattern. The company already operates Ivo Collaborate, a contract review platform used by Fortune 500 legal teams, and Sage appears designed to deepen that enterprise footprint by demonstrating technical credibility in a market growing skeptical of opaque AI claims. The broader industry context matters: as large language model providers race to offer vertical fine-tuning, legal operations teams have become a priority target, with contract management identified as the workflow most amenable to near-term automation.

Sage scored 91 percent on the Legal Agent Benchmark Contracts, an external evaluation that carries more weight than proprietary claims. On Ivo's own Ivo-micro1 Contract Bench — a five-dimension internal framework — the model correctly accepted 82 percent of counterparty changes and adhered to playbook standard-position criteria 51 percent of the time. The escalation metric is the headline caveat: Sage meets attorney escalation criteria just 23 percent of the time, meaning the model frequently misses signals that a human lawyer would flag for review. Performance on generating original contract points is also described as limited. Ivo says full benchmark results will be published in the coming weeks, a disclosure commitment that will be closely watched by enterprise procurement teams.

Industry Implications

Open-sourcing a legal AI model raises competitive pressure on incumbents who have kept their contract review logic proprietary, effectively forcing a conversation about what transparency in legal technology actually means for enterprise buyers. For legal operations teams, the availability of a benchmarked, inspectable model lowers the barrier to internal validation and potentially to customization against firm-specific playbooks. The 23 percent escalation rate is a data point competitors will cite, but it also sets a public baseline — one Ivo can improve against in full view of the market. More broadly, Sage reinforces the thesis that domain-specific fine-tuning of commodity foundation models, rather than training from scratch, is becoming the dominant architecture choice for professional services AI.

DreamLegal Perspective

Legal operations leaders evaluating contract AI should treat the pending full benchmark disclosure as a procurement trigger — not to buy, but to build an evaluation framework now. Sage's published metrics give legal tech buyers rare empirical ground to stand on when pressing other vendors for comparable transparency. The 23 percent escalation accuracy is the number to probe: any enterprise deploying AI-assisted contract review without a clear human-in-the-loop protocol for escalation decisions is carrying unquantified liability exposure. Watch whether competitors respond by releasing their own benchmark data or accelerating open-source plays of their own; either outcome reshapes how enterprise legal teams set baseline expectations for AI contract tools through 2027.