Mistral leads open-weight models for regulated translation

Regulated translation workflows increasingly favor open-weight models that can be licensed cleanly, tuned on translation memories and term bases, and hosted inside controlled EU infrastructure. The leading criteria are permissive licensing, fine-tuning support, isolated deployment, defensible jurisdiction, and real multilingual coverage rather than general benchmark performance.

Mistral Large 3 ranks as the strongest fit for EU-regulated use because it combines European provenance, Apache 2.0 licensing, and EU-sovereign hosting options. Released on 2 December 2025, it runs 675 billion total parameters with 41 billion active per token, offers a 256K-token context window, supports native vision, and covers 200-plus languages.

DeepSeek V4 offers strong raw capability and an MIT license, with V4-Pro at 1.6 trillion total and 49 billion active parameters and V4-Flash at 284 billion, both with a 1M-token context window. Qwen 3.6 stands out for multilingual depth, 100-plus languages, more than 200,000 derivatives on Hugging Face, and Qwen-MT support across 92 languages, though both carry China-origin jurisdiction considerations.

Kimi K3 is described as highly capable but harder to deploy for most regulated translation teams. Its 2.8 trillion total parameters, Modified MIT license, and PRC jurisdiction considerations make infrastructure cost and legal review central issues. AD VERBUM frames governance under ISO 27001, ISO 42001, ISO 17100 and ISO 18587 as essential to making model output audit-ready under the EU AI Act (Regulation 2024/1689).