Mistral Small 4 leads 3090 translation benchmarks

Mistral Small 4's 24B architecture proves that "sovereign" training beats brute-force parameter counts for European linguistic tasks.

• –Optimized MoE structure allows flagship performance to fit comfortably on a single RTX 3090 or 4090
• –Native French reasoning (Magistral core) eliminates the "translationese" artifacts common in English-centric models like Llama 4
• –Apache 2.0 licensing secures its position as the standard for private, local-first enterprise translation pipelines
• –Unified multimodal support (vision, code, translation) reduces VRAM overhead for complex automation workflows
• –Significant context window expansion (128k+) enables full-document consistency that traditional engines like DeepL still struggle to match