Mistral Releases 4-Model Apache 2.0 Family: 675B Sparse MoE Flagship + 3B Model Runs In-Browser — HSBC Signs Multi-Year Deal

Decrypt·medium signal

Mistral released four open-source models under Apache 2.0: Mistral Large 3 (675B total parameters, 41B active via sparse MoE, ranked #2 among open-source non-reasoning models on LMArena) and Ministral variants at 14B, 8B, and 3B — the 3B model runs in-browser via WebGPU at ~385 tokens/second on RTX 5090. HSBC announced a multi-year self-hosted Mistral deployment, citing GDPR and data sovereignty requirements as the primary driver. The release continues open-source competitive pressure on proprietary frontier models for regulated-industry workloads.