Ivo releases Sage, an open source model trained for long-horizon contract work

Contract intelligence company Ivo has released Sage, an open source AI model post-trained for long-horizon contract work, available for download from Hugging Face under an MIT license, which in general allows commercial use, modification and redistribution, subject to the terms of the base model. According to Ivo, it is the first open model from a legal AI company post-trained for this kind of work. The announcement was made on October 1 at its Inscribe customer conference and is detailed on the company's blog.

The starting point is a critique of generalist models: contract work is long, depends on context, chains tasks together and rarely has a single correct answer. Ivo argues that the least understood lever of legal AI is post-training, teaching the model the specific craft, and compares it to the difference between a fresh graduate and a lawyer who has already spent years in their specialty.

How it was built

Sage starts from DeepSeek V4 Flash, a public open model, and was post-trained with reinforcement learning on the Legal Agent Benchmark (LAB), a public set of long, multi-step legal tasks. Training used 1,040 tasks, of which 180 were fully reserved for the final evaluation, without the model seeing them during training; in addition, tasks based on the same agreement were kept together, all in or all out, to avoid leakage. The training infrastructure was run by its partner River AI. The company stresses that it used no customer data at any point.

What the results claim

The numbers are Ivo's and should be read as such. On contract tasks, Sage met 91% of the evaluation criteria, versus 70% for the base model, at a cost of 1.24 dollars per task; the frontier model it is compared against, Fable 5.1, cost 142.39 dollars on the same battery. On general legal tasks, its strict pass rate (15.2%) came in level with Fable 5.1 (15.7%), at 0.40 dollars per task versus 92.46. Post-training also changed how it works: 25 turns on average per task versus 31 for the base model, with 21% fewer tokens. Outside its own territory, it rose on RedlineBench (from 30.8 to 45.7) and stayed flat on LegalBench (83.0 versus 83.1 for the base), a sign that specializing did not degrade general legal reasoning.

Alongside the model, Ivo publishes the training method and the evaluation log: task lists, transcripts, deliverables and judge decisions. Its argument is that progress in contract AI depends on shared work, that transparency allows its claims to be audited and that every legal team can fine-tune an open model to its own data and safeguards.

Fit in the market

Most legal AI is closed. Makers such as Legora, which this month launched Skills to teach the agent each team's know-how, sell the result packaged in their platform. Ivo opens the opposite door: that a law firm or a department downloads the model, fine-tunes it with its data and governs it on its own, a philosophy similar to that of LawVu Lens, which reasons over the client's own contract portfolio. Both start from the same idea: the value lies in each organization's context. Anyone following the money will find the market context for these bets in our analysis of legaltech funding in 2026.

This story is part of our legaltech startups and AI section, where we track launches, rounds and deals in the sector.

Sources and limits

The launch, the architecture, the license and all performance and cost figures appear in Ivo's official announcement and are the maker's own data, not an independent evaluation. The comparison with other models uses the names and list prices Ivo publishes in its benchmark. The claim of being the first open model from a legal AI company post-trained for contracts is the company's own.