Analyse de l'application Fluent AI: Offline & Cloud LLM pour le 1 octobre

🤖 Fluent AI — Private Offline LLM + Claude, GPT-4 & Gemini Run AI entirely on your device — no cloud, no account, no data sent anywhere. Then switch to Claude, GPT-4 or Gemini when you need more power. One app. Every AI. Always private. ✨ WHAT'S NEW IN v1.3 🏥 MEDICAL AI (MedGemma) • Google's MedGemma 4B — clinical Q&A and biomedical text, 100% on-device • Requires accepting Google's Health AI Developer Foundation Terms • Not a substitute for professional medical advice 🤖 AGENTIC MODE • On-device AI agent with 12 built-in skills • Runs tasks autonomously: calendar events, web research, document digest, trip planning • Agent Task Inspector — see every reasoning step in real time • 3 free agent runs/day — no subscription needed to start • Scheduled tasks available with Premium ⚡ LITERT MTP — UP TO 2× FASTER • Gemma 4n E2B/E4B with Multi-Token Prediction on Android GPU • Speculative decoding — more tokens per step, same quality • Tok/s display measures decode-phase speed only for accurate results 👁️ ON-DEVICE VISION (Android) • Attach photos using Gemma 4n — processed entirely on-device • No image uploaded to any server, ever 🔒 PRIVACY FIRST • Conversations stay on your device • Optional local models = zero cloud data • API keys encrypted with AES — never stored in plain text • No mandatory account required 🧠 LOCAL AI MODELS • GGUF / llama.cpp: Gemma 3/4, Qwen 3.5, Phi-4, Llama, DeepSeek R1, Nemotron, MedGemma • LiteRT (Android GPU/NPU): Gemma 4n E2B/E4B — vision + MTP speculative decoding • Apple MLX: Native Metal on Apple Silicon and iOS 18+ (A17 Pro+) • Q5_K GPU acceleration on Qualcomm Adreno (alongside Q4_0) • Device-aware model recommendations based on your RAM and chipset • Browse, download, and manage models in-app — no sideloading needed • Import custom GGUF from HuggingFace URL or device storage ☁️ CLOUD AI • Claude (Anthropic), GPT-4 (OpenAI), Gemini (Google) • OpenRouter — 200+ models via a single API key • Streaming, vision, and tool calling across all providers 🌐 ONLINE SERVERS • Ollama Cloud and self-hosted Ollama • LM Studio, vLLM, LocalAI, and any OpenAI-compatible /v1 API • Multiple server profiles with per-profile encrypted auth headers 🎤 VOICE MODE • 5 conversation modes: Normal, Interview, Learning, Storytelling, Translation • Animated waveform, voice commands (speed, repeat, stop) • Quick-capture mic button directly in the chat input bar 📚 KNOWLEDGE BASES (RAG) • Import PDFs, TXT, and Markdown — AI references your docs when answering • Semantic search for relevant context, topic and project organisation 🔧 POWER FEATURES • Tool calling: Calculator, DateTime, Weather, Web Search, mem0 Memory • MCP servers: GitHub, Slack, Notion, Supabase, and 20+ presets • Code execution: Python, Bash, Node.js from code blocks (desktop + mobile JS) • Model benchmarking: tok/s, TTFT, MMLU-50 quality score, shareable PNG cards • Slash commands: /agent, /clear, /export, /voice, /template and more • Per-chat thinking toggle for Qwen3, DeepSeek R1, Nemotron reasoning models • URL context injection — paste a link, AI reads the page for context • Polish Before Send — AI rewrites your draft before you hit send • Continue button — resumes responses cut off at the token limit 📁 CHAT ORGANISATION • Folders, tags, and cross-chat full-text search across every message • HuggingFace model browser with bookmarks and memory fitness badges • Conversation branching and message reactions 🌟 PREMIUM (OPTIONAL) • Ad-free experience • Scheduled agent tasks (recurring or one-time) • Priority feature access and advanced analytics 📱 PERFECT FOR ✓ Privacy-focused users — local models, zero cloud data ✓ Android power users — LiteRT GPU/NPU with MTP acceleration ✓ Developers — benchmark GGUF, LiteRT, and MLX side-by-side ✓ Healthcare researchers — MedGemma on-device, no upload needed ✓ Students — knowledge bases for study documents and materials ✓ Professionals — agentic tasks, document Q&A, and tool calling