{ "metadata": { "generated_at": "2026-05-26T10:18:00+02:00", "project": "Can My GPU Run This LLM", "dataset_type": "local_ai_applications", "application_count": 7, "data_policy": "Curated application scenarios for planning. Requirements are practical estimates, not benchmarks.", "update_cadence": "Review weekly, and immediately when a major coding agent, local model, Ollama feature, or workflow platform release changes the hardware or workflow assumptions.", "last_curated_by": "Codex", "watchlist": [ { "name": "OpenAI Codex", "url": "https://openai.com/index/introducing-codex/", "watch_for": "New Codex app, CLI, SDK, automations, workspace-agent, or coding-model releases" }, { "name": "Anthropic Claude Code", "url": "https://www.anthropic.com/product/claude-code", "watch_for": "Agentic coding workflow, CLI, IDE, and multi-agent orchestration changes" }, { "name": "Google Gemini CLI", "url": "https://developers.google.com/gemini-code-assist/docs/gemini-cli", "watch_for": "Terminal-agent, ReAct, MCP, GitHub Actions, and Code Assist agent-mode changes" }, { "name": "Ollama", "url": "https://ollama.com/blog/tool-support", "watch_for": "Tool calling, local runtime, model-library, and agent-enabling updates" } ], "candidate_rules": [ "Add a new application when an official release enables a new repeatable workflow, not just a model rename.", "Prefer official sources, docs, release notes, or first-party blogs before adding claims.", "Mark private Wolfgang stacks as reference setups, not public market trends.", "Never let partner links change hardware fit, application ranking, or cloud fallback logic." ] }, "applications": [ { "id": "hermes-kellerrechner-agent", "name": "Hermes / Kellerrechner Agent", "category": "Personal AI assistant", "type": "private_reference_stack", "trend_stage": "active internal reference", "local_first": true, "minimum_vram_gb": 12, "recommended_vram_gb": 24, "recommended_system_ram_gb": 32, "recommended_models": ["qwen2-5-coder-14b", "qwen3-8b", "llama-3-1-8b"], "components": ["Ollama", "Telegram", "n8n", "Obsidian", "Control Center"], "best_for": "A local operator that answers through Telegram, maintains project context, and supports technical tasks.", "hardware_note": "12GB VRAM is a workable local-agent floor. 24GB VRAM gives better headroom for coding, long context, and parallel tasks.", "cloud_fallback": true, "update_signals": ["Ollama tool support", "new Qwen coder models", "Telegram gateway stability", "Hermes role changes"] }, { "id": "gravity-claw-business-operator", "name": "Gravity Claw Business Operator", "category": "Business automation agent", "type": "private_reference_stack", "trend_stage": "active internal reference", "local_first": false, "minimum_vram_gb": 12, "recommended_vram_gb": 24, "recommended_system_ram_gb": 64, "recommended_models": ["qwen2-5-coder-14b", "qwen2-5-coder-32b", "llama-3-1-8b"], "components": ["Control Center", "Operator Inbox", "n8n", "Obsidian", "Social approval queue"], "best_for": "Monitoring projects, tracking next steps, preparing approved updates, and keeping a one-person business system coherent.", "hardware_note": "Hybrid is best: local models for routine routing and summaries, cloud/API models for heavy research and polished public drafts.", "cloud_fallback": true, "update_signals": ["workspace agents", "Codex automations", "MCP integrations", "business workflow agents"] }, { "id": "local-coding-agent", "name": "Local Coding Agent", "category": "Coding agent", "type": "public_trend", "trend_stage": "fast-moving", "local_first": true, "minimum_vram_gb": 12, "recommended_vram_gb": 24, "recommended_system_ram_gb": 64, "recommended_models": ["qwen2-5-coder-14b", "qwen2-5-coder-32b", "phi-4-14b"], "components": ["Terminal", "Git", "test runner", "repo context", "approval workflow"], "best_for": "Local code Q&A, scripts, small repo edits, review support, and supervised terminal workflows.", "hardware_note": "12GB can run strong 7B/14B coder models. 24GB is the practical workstation tier for heavier local coding agents.", "cloud_fallback": true, "update_signals": ["Codex", "Claude Code", "Gemini CLI", "Qwen coder releases", "local tool calling"] }, { "id": "telegram-ai-bot", "name": "Telegram AI Bot", "category": "Chat and automation bot", "type": "repeatable_workflow", "trend_stage": "stable", "local_first": true, "minimum_vram_gb": 8, "recommended_vram_gb": 12, "recommended_system_ram_gb": 32, "recommended_models": ["qwen3-8b", "llama-3-1-8b", "mistral-7b"], "components": ["Telegram Bot API", "Ollama", "SQLite", "webhook or polling worker"], "best_for": "Personal commands, status checks, reminders, approval buttons, and lightweight agent routing.", "hardware_note": "8GB is enough for small chat bots. 12GB is safer once tools, memory, and longer prompts are added.", "cloud_fallback": false, "update_signals": ["Telegram Bot API changes", "small model releases", "tool-calling support", "gateway errors"] }, { "id": "knowledge-vault-agent", "name": "Obsidian / Knowledge Vault Agent", "category": "Knowledge and memory agent", "type": "repeatable_workflow", "trend_stage": "growing", "local_first": true, "minimum_vram_gb": 12, "recommended_vram_gb": 24, "recommended_system_ram_gb": 64, "recommended_models": ["qwen2-5-coder-14b", "llama-3-1-8b", "qwen3-8b"], "components": ["Obsidian", "local files", "embeddings", "RAG", "note writer"], "best_for": "Private project memory, structured notes, document summaries, and searchable local knowledge.", "hardware_note": "Small models work for tagging and summaries. Long documents and RAG benefit from 24GB VRAM or cloud fallback.", "cloud_fallback": true, "update_signals": ["RAG libraries", "embedding models", "Obsidian workflow changes", "local long-context models"] }, { "id": "social-publishing-approval-worker", "name": "Social Publishing Approval Worker", "category": "Marketing automation agent", "type": "private_reference_stack", "trend_stage": "active internal reference", "local_first": false, "minimum_vram_gb": 8, "recommended_vram_gb": 12, "recommended_system_ram_gb": 32, "recommended_models": ["qwen3-8b", "llama-3-1-8b", "qwen2-5-coder-7b"], "components": ["Diff watcher", "Telegram approval", "Bluesky API", "Mastodon API", "SQLite queue"], "best_for": "Turning data changes into reviewed news-style drafts before anything is published externally.", "hardware_note": "Local draft triage can run on 8GB-12GB. Public copy quality may still benefit from a stronger cloud model.", "cloud_fallback": true, "update_signals": ["model price diffs", "agent releases", "Bluesky/Mastodon API changes", "approval queue errors"] }, { "id": "desktop-multi-agent-command-center", "name": "Desktop Multi-Agent Command Center", "category": "Agent orchestration", "type": "public_trend", "trend_stage": "fast-moving", "local_first": false, "minimum_vram_gb": 24, "recommended_vram_gb": 48, "recommended_system_ram_gb": 128, "recommended_models": ["qwen2-5-coder-32b", "llama-3-1-70b", "deepseek-r1-distill-qwen-32b"], "components": ["multiple agents", "sandboxes", "worktrees", "approval UI", "logs"], "best_for": "Running several coding, research, or operator agents in parallel with human supervision.", "hardware_note": "This is the workstation tier. Consumer 24GB GPUs can help, but parallel local agents quickly push toward 48GB+ or cloud.", "cloud_fallback": true, "update_signals": ["Codex app", "Claude Code", "Gemini CLI", "workspace agents", "multi-agent workflow releases"] } ] }