Mistral Large 3 ships as a 675B-parameter MoE under a true Apache 2.0 license
Mistral Large 3 arrives as a sparse mixture-of-experts model — 41 billion active parameters, 675 billion total, a 256K context window — released under a genuine Apache 2.0 license. A frontier-scale European open-weight model with permissive terms puts real pressure on the closed labs and on the provenance question for enterprises.
The architecture is built for deployable scale. A 675-billion-parameter total with only 41 billion active per token is the mixture-of-experts bargain — frontier-scale capacity while keeping inference tractable — and shipping it with a 256K context makes it viable for the long-document and agentic workloads that bounded context otherwise blocks. This is a serious model, not a research artifact.
The Apache 2.0 license is the strategic core. A true permissive license lets enterprises modify, self-host, and build products without the restrictions that made some earlier 'open' releases evaluation-only. For a European lab, it is also a provenance argument: a Western, Apache-licensed frontier open model is deployable in ways that address the risk-posture concerns some enterprises have about the leading Chinese open weights.
The competitive pressure is on both closed labs and the open field. A 675B open MoE narrows what the closed premium buys, and it gives the open ecosystem a Western flagship to sit alongside Qwen and DeepSeek at the top tier. As the EU's training-data-summary duties become enforceable, a lab shipping under clear terms with clear provenance is also positioning for the compliance environment the closed labs must navigate too.
TechJack Solutions — Best open-source AI models 2026: 5 that rival frontier → · Hugging Face — Best open-source LLM models in 2026 →