Ivo Sage Open-Source Contract Intelligence Model
The first open-source model from a legal AI company, post-trained for long-horizon contract work with MIT-licensed weights
Tool Overview
Details, capabilities and usage guidance
Features
- ✓ Free to download: MIT-licensed weights published on Hugging Face, self-hostable
- ✓ Post-trained on DeepSeek V4 Flash with reinforcement learning; LAB Contracts criterion pass rate rose from 70% to 91%
- ✓ Fine-tunable on your own contract data, with rules and safeguards you can adapt to specific use cases
- ✓ Public evaluations and release details so developers can verify and reproduce
- ✓ Comes with the Ivo-micro1 Contract Bench, scoring five dimensions of attorney judgement: prioritization, restraint, deal-specific adaptation, escalation and playbook adherence
How to Use
- Open the Hugging Face model page ivo-ai-labs/Ivo-Sage and read the model card and MIT license
- Download the weights and run them locally or in your own environment
- Fine-tune on your own contracts and playbooks, then adjust rules and safeguards
- Check performance against the published evaluations and the Ivo-micro1 Contract Bench before wiring it into production
FAQ
What is Ivo Sage?
An open-source model Ivo announced on 1 October 2026 at Ivo Inscribe, its first user conference, in San Francisco. Ivo says it is the first model released by a legal AI company that is post-trained for long-horizon contract work, and that attorneys, researchers and developers can download, run and build on it.
What license does it use, and is it really free?
Ivo released MIT-licensed adapter weights, hosted on Hugging Face at ivo-ai-labs/Ivo-Sage. Ivo's statement is that anyone can download and run it for free; you still need your own inference compute, and that cost sits with you.
Is it trained from scratch?
No. Per the release, it is a set of fine-tuned weights layered on DeepSeek V4 Flash, built in partnership with River AI by post-training on long-horizon contract work using public and synthetic data generated by real attorneys, followed by reinforcement learning.
How much better is it?
Ivo says that after reinforcement learning the model went from meeting 70% of the pass criteria on the Legal Agent Benchmark (LAB) Contracts to 91%, reaching comparable quality to much larger frontier models with higher token efficiency at a fraction of the cost.
Can it replace attorneys for contract review?
No. Ivo's own published findings say the hardest judgement happens before any edit: models got the right call only 46% of the time when the correct response was to counter or reject, and met only 23% of expert escalation criteria on average, so knowing when to hand a call to a human remains a gap.
What else was announced?
Ivo previewed a new contract review benchmark, the Ivo-micro1 Contract Bench, built with data lab micro1 to measure review judgement, and announced general availability of Ivo Collaborate, covering the contract lifecycle from intake through signature.