mistral-performance-tuning

# Mistral AI Performance Tuning ## Overview Optimize Mistral AI API response times and throughput. Key levers: model selection (Mistral Small ~200ms TTFT vs Large ~500ms), prompt length (fewer tokens = faster), streaming (perceived speed), caching (zero-latency repeats), and concurrent request management. ## Prerequisites - Mistral API integration in production - Understanding of RPM/TPM limits for your tier - Application architecture supporting streaming ## Instructions ### Step 1: Model Selection by Latency Budget ```typescript const MODELS_BY_USE_CASE: Record<string, { model: string; ttftMs: string; note: string }> = { realtime_chat: { model: 'mistral-small-latest', ttftMs: '~200ms', note: '256k ctx, cheapest' }, code_completion: { model: 'codestral-latest', ttftMs: '~150ms', note: 'Optimized for code + FIM' }, code_agents: { model: 'devstral-latest', ttftMs: '~300ms', note: 'Agentic coding tasks' }, reasoning: { model: 'mistral-large-latest', ttftMs: '~500ms', note: '256k ctx, strongest' }, vision: { model: 'pixtral-large-latest', ttftMs: '~600ms', note: 'Image + text multimodal' }, embeddings: { model: 'mistral-embed', ttftMs: '~50ms', note: '1024-dim, batch-friendly' }, edge_devices: { model: 'ministral-latest', ttftMs: '~100ms', note: '3B-14B, fastest' }, }; ``` ### Step 2: Streaming for User-Facing Responses Streaming reduces perceived latency from 1-2s (full response) t...