Australian-made LLM beats OpenAI and Google at legal retrieval
Isaacus, an Australian foundational legal AI startup, has launched Kanon 2 Embedder, a state-of-the-art legal embedding LLM, and unveiled the Massive Legal Embedding Benchmark (MLEB), an open-source benchmark for evaluating legal information retrieval performance across six jurisdictions (the US, UK, EU, Australia, Singapore, and Ireland) and five domains (cases, statutes, regulations, contracts, and academia).
Kanon 2 Embedder ranks first on MLEB as of 23 October 2025, delivering 9% higher accuracy than OpenAI Text Embedding 3 Large and 6% higher accuracy than Google Gemini Embedding while running >30% faster than both LLMs. Kanon 2 Embedder leads a field of 20 LLMs, including Qwen3 Embedding 8B, IBM Granite Embedding R2, and Microsoft E5 Large Instruct.
The quality of search results sets the ceiling for legal RAG applications", said Umar Butler, founder of Isaacus. "Kanon 2 Embedder raises that ceiling, dramatically.
Why it matters
Embedding models convert documents and queries into sets of numbers known as ‘embeddings’ that can be compared with each other to identify relevant search results.
Embeddings power the ‘retrieval’ component of retrieval-augmented generation (RAG) applications and are widely used across the legal tech industry, including by Harvey, LexisNexis, and Relativity.
In legal RAG applications, low-quality embeddings lead to low-quality search results, which in turn lead to low-quality responses and increased hallucinations.
Despite their importance, limited attention has been paid to ensuring that embedding models are genuinely fit for legal information retrieval.
A new gold standard for legal information retrieval
Isaacus believes MLEB is the most comprehensive and diverse legal embedding benchmark to date. It spans multiple countries, document types, and areas of law. Every dataset in MLEB has been curated and vetted by domain experts.
On MLEB, legal-tuned LLMs consistently outperform similar-sized general-purpose LLMs. The top-ranked Kanon 2 Embedder is derived from Kanon 2, a new legal foundation model trained on millions of laws, regulations, cases, contracts, and papers from 38 jurisdictions. The second- and third-best LLMs—Voyage 3 Large and Voyage 3.5—were also optimized for law, possibly thanks to Voyage’s (owned by MongoDB) partnership with Harvey.
Key achievements
- Kanon 2 Embedder ranks first on MLEB while running 340% faster and being several times smaller than the second-ranked LLM, Voyage 3 Large.
- MLEB covers decisions, laws, regulations, contracts, and textbooks across the US, UK, Australia, Singapore, and Ireland.
Isaacus hopes Kanon 2 Embedder and MLEB will elevate the quality of legal retrieval globally by setting a new gold standard for what superior legal search capabilities and benchmarks look like.
Respecting legal data sovereignty
Isaacus takes the sensitivity of legal data seriously. Unlike others in the embeddings space (see, for example, Voyage, Cohere, and Jina), Isaacus has chosen to not opt businesses into using their private data for training by default.
Additionally, for customers with heightened privacy, security, or reliability concerns, Isaacus will soon be offering air-gapped model deployment containers on AWS Marketplace and Microsoft Marketplace.
Available today
Isaacus is inviting the legal tech industry to try Kanon 2 Embedder for themselves and see whether it delivers superior retrieval results for their own use cases. They can do so by following the quick start guide on the Isaacus docs.
Enterprises interested in private AWS and Azure deployments should follow Isaacus on LinkedIn to get updated in the next few weeks on the release of their AWS Marketplace and Microsoft Marketplace products.
Isaacus has also made the data and code behind MLEB openly and freely available on Hugging Face and GitHub. The MLEB leaderboard is hosted on their website, as is a full writeup of their methodology, soon to be released as a paper.
