Australian-made LLM beats OpenAI and Google at legal retrieval (huggingface.co)

Isaacus, an Australian legal AI startup, released Kanon 2 Embedder — a legal-specialized embedding LLM — alongside the Massive Legal Embedding Benchmark (MLEB), an open-source evaluation suite covering six jurisdictions (US, UK, EU, Australia, Singapore, Ireland) and five document domains (cases, statutes, regulations, contracts, academia/textbooks). On MLEB (as of 23 Oct 2025) Kanon 2 Embedder ranked first out of 20 models, scoring 9% higher accuracy than OpenAI Text Embedding 3 Large and 6% higher than Google Gemini Embedding while running >30% faster than both; Isaacus also reports it is 340% faster and several times smaller than the second-ranked Voyage 3 Large. Kanon 2 is derived from a legal foundation model trained on millions of laws, regulations, cases, contracts and papers from 38 jurisdictions, and MLEB’s datasets were curated and vetted by domain experts. The win matters because embeddings are the retrieval backbone of legal RAG systems: higher-quality embeddings reduce bad search hits and downstream hallucinations. MLEB addresses a long-standing gap by providing a diverse, transparent benchmark that shows law-tuned models outperform similar-sized general-purpose LLMs. Isaacus has open-sourced MLEB’s data and code (Hugging Face, GitHub), published a leaderboard and methodology, and emphasizes privacy — opting users out of data-for-training by default and planning air-gapped AWS/Azure deployment containers for sensitive legal workloads.