Julius AI launches Deep Finance benchmark and finds Thinking Machines' open-source model tops American AI for finance tasks

Julius AI's Deep Finance benchmark

Julius AI has released Deep Finance, what it claims is the first long-horizon benchmark designed to evaluate AI models on real investment banking and private equity deal workflows. The benchmark was built with help from professionals with backgrounds at Evercore, JPMorgan, UBS, and Harvard Business School.

The core test is deliberately open-ended: models are given a messy data room, raw files, and a high-level instruction to assess whether a deal is worth pursuing. Sonwalkar says the key distinction is whether models actually trace revenue figures down to source documents and reconcile discrepancies, rather than accepting a headline number. In real deals, he notes, management reports, audit reports, and independently calculated figures frequently diverge. Models that perform well in coding tasks tend to pattern-match by searching for a keyword and stopping there, which doesn't hold up in finance.

Julius ran the benchmark using each model's native CLI rather than its own harness, so the results reflect Codex for OpenAI models and Claude Code for Anthropic's, keeping the evaluation independent.

“We just released Deep Finance — the first ever long horizon high finance benchmark. We work with experts from firms like Evercore, JPMorgan, UBS, and Harvard Business School to build this benchmark. When we ran Thinking Machines' InkLink 2 Small on our eval, we found that it's actually the best American open source model and second best open source model overall, after Kimi.”

Thinking Machines result

Thinking Machines released InkLink 2 Small the same day Julius published its benchmark, without access to the eval during training. When Julius ran it, InkLink 2 Small ranked as the best American open-source model for finance tasks and second overall, behind Kimi. It also outperformed NVIDIA's Nemotron. Sonwalkar sees the jump from InkLink 1 to InkLink 2 as meaningful evidence that the company's approach is working. Thinking Machines is separately reported to be closing a funding round at a $5 billion raise on a $40 billion valuation, which Sonwalkar describes as impressive. The company is also said to be a significant partner to Reality's fund, and Sonwalkar suggests that financial partners are likely influencing the model's capabilities in ways that show up even in the open-source release.

What Julius does with this

Sonwalkar is candid that the commercial path isn't fully defined. Julius Research, the new division behind the benchmark, was set up to pursue problems relevant to knowledge work broadly, not just to build a sales asset. That said, the benchmark creates optionality: potential fine-tuning partnerships with model labs, integration into Julius's own product, or use as a training signal for financial institutions that want to run models on their own infrastructure. The open-source angle matters here because data sovereignty is a real constraint for financial institutions and government agencies, and a strong open-source baseline they can further fine-tune addresses that directly.

The gap in American open-source model quality for finance is the pressure point Julius is measuring. InkLink 2 Small's result suggests that gap is closing faster than the field expected.