{ "metadata": { "generated_at": "2026-05-27T10:20:04.960Z", "project": "Can My GPU Run This LLM", "dataset_type": "agent_model_guide", "model_count": 23, "source_files": [ "/data/models.json", "/data/applications.json" ], "policy": "Agent-facing planning data. Estimates are practical guidance, not benchmarks.", "commercial_policy": "Technical recommendations and compatibility status must not be changed by affiliate, referral, sponsorship, or discount relationships.", "ranking_influenced_by_affiliate": false, "last_curated_by": "Codex" }, "context_profiles": [ { "id": "quick", "label": "Quick chat / small prompt", "multiplier": 0.9, "note": "Short Q&A, shell help, small config snippets, quick translation." }, { "id": "coding", "label": "Coding assistant / scripts", "multiplier": 1, "note": "Focused coding, small repo edits, review support, debugging one or two files." }, { "id": "repo", "label": "Repo or long chat", "multiplier": 1.15, "note": "Longer conversations, README plus source files, multi-step code reasoning." }, { "id": "rag", "label": "PDF / document analysis", "multiplier": 1.35, "note": "Document summaries, meeting notes, research pages, RAG-style retrieval prompts." }, { "id": "agent", "label": "Agent with tools", "multiplier": 1.45, "note": "Tool calls, planning loops, repeated instructions, memory, and workflow state." }, { "id": "extreme", "label": "Extreme long context", "multiplier": 1.7, "note": "Large document batches, whole-project context, heavy RAG, or long autonomous sessions." } ], "quantization_profiles": [ { "id": "q4", "bits": 4 }, { "id": "q5", "bits": 5 }, { "id": "q8", "bits": 8 }, { "id": "fp16", "bits": 16 } ], "models": [ { "id": "llama-3-1-8b", "model_name": "Llama 3.1 8B Instruct", "family": "Llama", "provider": "Meta", "license_type": "open-weight", "parameters_billions": 8, "architecture": "dense", "ollama_tag": "llama3.1:8b", "recommended_quantization": "Q4_K_M", "best_for": [ "chat", "agents/tool workflows", "Fast local chat, lightweight agents, low-cost local testing" ], "weak_for": [ "coding", "vision/image understanding", "reasoning", "Complex coding and deep reasoning" ], "not_for": [ "image understanding / multimodal vision", "coding-heavy workloads", "deep multi-step reasoning" ], "agent_readiness_score": 78, "context": { "minimum_context_tokens": 8192, "typical_context_tokens": 128000, "context_fit_note": "Large context-capable on paper; verify runtime memory and backend limits." }, "memory_estimates_gb": { "quantization": { "q4": 6, "q5": 7.5, "q8": 12, "fp16": 24 }, "q4_context_profiles": { "quick": 5.4, "coding": 6, "repo": 6.9, "rag": 8.1, "agent": 8.7, "extreme": 10.2 }, "caveat": "Estimates are practical planning estimates. Real use varies by backend, KV cache, context length, drivers, quantization file, and offloading." }, "hardware": { "minimum_vram_gb_q4": 8, "comfortable_vram_gb_q4": 12, "recommended_gpu": "RTX 4060 8GB or better", "local_fit_note": "Good small-local-model candidate for 8GB+ GPUs at Q4." }, "routing_guidance": { "green": "Prefer local inference first if the selected GPU has comfortable headroom.", "yellow": "Reduce context, use lower quantization, try a smaller model, or expect RAM/offload slowdown.", "red": "Use a smaller local model, larger GPU, cloud GPU fallback, or compare API/cloud costs.", "commercial_option_ids": [], "ranking_influenced_by_affiliate": false }, "recommended_applications": [ { "id": "hermes-kellerrechner-agent", "name": "Hermes / Kellerrechner Agent", "category": "Personal AI assistant" }, { "id": "gravity-claw-business-operator", "name": "Gravity Claw Business Operator", "category": "Business automation agent" }, { "id": "telegram-ai-bot", "name": "Telegram AI Bot", "category": "Chat and automation bot" }, { "id": "knowledge-vault-agent", "name": "Obsidian / Knowledge Vault Agent", "category": "Knowledge and memory agent" }, { "id": "social-publishing-approval-worker", "name": "Social Publishing Approval Worker", "category": "Marketing automation agent" } ], "sources": [ { "label": "Ollama library", "url": "https://ollama.com/library/llama3.1" } ], "last_verified": "2026-05-19", "confidence": "high" }, { "id": "llama-3-1-70b", "model_name": "Llama 3.1 70B Instruct", "family": "Llama", "provider": "Meta", "license_type": "open-weight", "parameters_billions": 70, "architecture": "dense", "ollama_tag": "llama3.1:70b", "recommended_quantization": "Q4_K_M", "best_for": [ "chat", "coding", "agents/tool workflows", "reasoning", "High-quality local chat and reasoning on workstation-class hardware" ], "weak_for": [ "vision/image understanding", "Too large for single 24GB consumer GPUs without heavy offload" ], "not_for": [ "image understanding / multimodal vision" ], "agent_readiness_score": 88, "context": { "minimum_context_tokens": 8192, "typical_context_tokens": 128000, "context_fit_note": "Large context-capable on paper; verify runtime memory and backend limits." }, "memory_estimates_gb": { "quantization": { "q4": 44, "q5": 55, "q8": 88, "fp16": 176 }, "q4_context_profiles": { "quick": 39.6, "coding": 44, "repo": 50.6, "rag": 59.4, "agent": 63.8, "extreme": 74.8 }, "caveat": "Estimates are practical planning estimates. Real use varies by backend, KV cache, context length, drivers, quantization file, and offloading." }, "hardware": { "minimum_vram_gb_q4": 48, "comfortable_vram_gb_q4": 64, "recommended_gpu": "48GB+ VRAM workstation or multi-GPU setup", "local_fit_note": "Large local model. Prefer 48GB+ VRAM, multi-GPU, cloud GPU, or hosted API fallback." }, "routing_guidance": { "green": "Prefer local inference first if the selected GPU has comfortable headroom.", "yellow": "Reduce context, use lower quantization, try a smaller model, or expect RAM/offload slowdown.", "red": "Use a smaller local model, larger GPU, cloud GPU fallback, or compare API/cloud costs.", "commercial_option_ids": [ "runpod-cloud-gpu-fallback", "apiroute-cloud-api-comparison" ], "ranking_influenced_by_affiliate": false }, "recommended_applications": [ { "id": "desktop-multi-agent-command-center", "name": "Desktop Multi-Agent Command Center", "category": "Agent orchestration" } ], "sources": [ { "label": "Ollama library", "url": "https://ollama.com/library/llama3.1" } ], "last_verified": "2026-05-19", "confidence": "high" }, { "id": "qwen2-5-coder-7b", "model_name": "Qwen2.5 Coder 7B", "family": "Qwen", "provider": "Alibaba", "license_type": "open-weight", "parameters_billions": 7, "architecture": "dense", "ollama_tag": "qwen2.5-coder:7b", "recommended_quantization": "Q4_K_M", "best_for": [ "chat", "coding", "agents/tool workflows", "Small local coding assistant and agent tool generation" ], "weak_for": [ "vision/image understanding", "reasoning", "Larger refactors and complex multi-file reasoning" ], "not_for": [ "image understanding / multimodal vision", "deep multi-step reasoning" ], "agent_readiness_score": 82, "context": { "minimum_context_tokens": 8192, "typical_context_tokens": 32768, "context_fit_note": "Good practical context range for coding, chat, and moderate RAG." }, "memory_estimates_gb": { "quantization": { "q4": 5.5, "q5": 6.88, "q8": 11, "fp16": 22 }, "q4_context_profiles": { "quick": 4.95, "coding": 5.5, "repo": 6.32, "rag": 7.43, "agent": 7.97, "extreme": 9.35 }, "caveat": "Estimates are practical planning estimates. Real use varies by backend, KV cache, context length, drivers, quantization file, and offloading." }, "hardware": { "minimum_vram_gb_q4": 8, "comfortable_vram_gb_q4": 12, "recommended_gpu": "RTX 4060 8GB or better", "local_fit_note": "Good small-local-model candidate for 8GB+ GPUs at Q4." }, "routing_guidance": { "green": "Prefer local inference first if the selected GPU has comfortable headroom.", "yellow": "Reduce context, use lower quantization, try a smaller model, or expect RAM/offload slowdown.", "red": "Use a smaller local model, larger GPU, cloud GPU fallback, or compare API/cloud costs.", "commercial_option_ids": [], "ranking_influenced_by_affiliate": false }, "recommended_applications": [ { "id": "social-publishing-approval-worker", "name": "Social Publishing Approval Worker", "category": "Marketing automation agent" } ], "sources": [ { "label": "Ollama library", "url": "https://ollama.com/library/qwen2.5-coder" } ], "last_verified": "2026-05-19", "confidence": "high" }, { "id": "qwen2-5-coder-14b", "model_name": "Qwen2.5 Coder 14B", "family": "Qwen", "provider": "Alibaba", "license_type": "open-weight", "parameters_billions": 14, "architecture": "dense", "ollama_tag": "qwen2.5-coder:14b", "recommended_quantization": "Q4_K_M", "best_for": [ "chat", "coding", "agents/tool workflows", "reasoning", "Local coding, scripts, repo assistance, technical agents" ], "weak_for": [ "vision/image understanding", "Can be tight on 12GB GPUs at longer context" ], "not_for": [ "image understanding / multimodal vision" ], "agent_readiness_score": 88, "context": { "minimum_context_tokens": 8192, "typical_context_tokens": 32768, "context_fit_note": "Good practical context range for coding, chat, and moderate RAG." }, "memory_estimates_gb": { "quantization": { "q4": 10.5, "q5": 13.13, "q8": 21, "fp16": 42 }, "q4_context_profiles": { "quick": 9.45, "coding": 10.5, "repo": 12.07, "rag": 14.18, "agent": 15.22, "extreme": 17.85 }, "caveat": "Estimates are practical planning estimates. Real use varies by backend, KV cache, context length, drivers, quantization file, and offloading." }, "hardware": { "minimum_vram_gb_q4": 12, "comfortable_vram_gb_q4": 16, "recommended_gpu": "RTX 3060 12GB minimum, 16GB+ preferred", "local_fit_note": "Practical 12GB local-agent candidate at Q4 with headroom checks." }, "routing_guidance": { "green": "Prefer local inference first if the selected GPU has comfortable headroom.", "yellow": "Reduce context, use lower quantization, try a smaller model, or expect RAM/offload slowdown.", "red": "Use a smaller local model, larger GPU, cloud GPU fallback, or compare API/cloud costs.", "commercial_option_ids": [], "ranking_influenced_by_affiliate": false }, "recommended_applications": [ { "id": "hermes-kellerrechner-agent", "name": "Hermes / Kellerrechner Agent", "category": "Personal AI assistant" }, { "id": "gravity-claw-business-operator", "name": "Gravity Claw Business Operator", "category": "Business automation agent" }, { "id": "local-coding-agent", "name": "Local Coding Agent", "category": "Coding agent" }, { "id": "knowledge-vault-agent", "name": "Obsidian / Knowledge Vault Agent", "category": "Knowledge and memory agent" } ], "sources": [ { "label": "Ollama library", "url": "https://ollama.com/library/qwen2.5-coder" } ], "last_verified": "2026-05-19", "confidence": "high" }, { "id": "qwen2-5-coder-32b", "model_name": "Qwen2.5 Coder 32B", "family": "Qwen", "provider": "Alibaba", "license_type": "open-weight", "parameters_billions": 32, "architecture": "dense", "ollama_tag": "qwen2.5-coder:32b", "recommended_quantization": "Q4_K_M", "best_for": [ "chat", "coding", "agents/tool workflows", "reasoning", "Strong local coding and architecture work on 24GB GPUs" ], "weak_for": [ "vision/image understanding", "Little VRAM headroom on single 24GB GPUs with long context" ], "not_for": [ "image understanding / multimodal vision" ], "agent_readiness_score": 92, "context": { "minimum_context_tokens": 8192, "typical_context_tokens": 32768, "context_fit_note": "Good practical context range for coding, chat, and moderate RAG." }, "memory_estimates_gb": { "quantization": { "q4": 21, "q5": 26.25, "q8": 42, "fp16": 84 }, "q4_context_profiles": { "quick": 18.9, "coding": 21, "repo": 24.15, "rag": 28.35, "agent": 30.45, "extreme": 35.7 }, "caveat": "Estimates are practical planning estimates. Real use varies by backend, KV cache, context length, drivers, quantization file, and offloading." }, "hardware": { "minimum_vram_gb_q4": 24, "comfortable_vram_gb_q4": 32, "recommended_gpu": "RTX 3090/4090 24GB minimum, 32GB+ preferred", "local_fit_note": "Workstation-local candidate. Prefer 32GB+ VRAM for agents or long context." }, "routing_guidance": { "green": "Prefer local inference first if the selected GPU has comfortable headroom.", "yellow": "Reduce context, use lower quantization, try a smaller model, or expect RAM/offload slowdown.", "red": "Use a smaller local model, larger GPU, cloud GPU fallback, or compare API/cloud costs.", "commercial_option_ids": [ "apiroute-cloud-api-comparison" ], "ranking_influenced_by_affiliate": false }, "recommended_applications": [ { "id": "gravity-claw-business-operator", "name": "Gravity Claw Business Operator", "category": "Business automation agent" }, { "id": "local-coding-agent", "name": "Local Coding Agent", "category": "Coding agent" }, { "id": "desktop-multi-agent-command-center", "name": "Desktop Multi-Agent Command Center", "category": "Agent orchestration" } ], "sources": [ { "label": "Ollama library", "url": "https://ollama.com/library/qwen2.5-coder" } ], "last_verified": "2026-05-19", "confidence": "high" }, { "id": "qwen3-8b", "model_name": "Qwen3 8B", "family": "Qwen", "provider": "Alibaba", "license_type": "open-weight", "parameters_billions": 8, "architecture": "dense", "ollama_tag": "qwen3:8b", "recommended_quantization": "Q4_K_M", "best_for": [ "chat", "coding", "agents/tool workflows", "reasoning", "Fast general local assistant with reasoning/coding balance" ], "weak_for": [ "vision/image understanding", "Less capable than 14B/32B models for large tasks" ], "not_for": [ "image understanding / multimodal vision" ], "agent_readiness_score": 86, "context": { "minimum_context_tokens": 8192, "typical_context_tokens": 40000, "context_fit_note": "Good practical context range for coding, chat, and moderate RAG." }, "memory_estimates_gb": { "quantization": { "q4": 6, "q5": 7.5, "q8": 12, "fp16": 24 }, "q4_context_profiles": { "quick": 5.4, "coding": 6, "repo": 6.9, "rag": 8.1, "agent": 8.7, "extreme": 10.2 }, "caveat": "Estimates are practical planning estimates. Real use varies by backend, KV cache, context length, drivers, quantization file, and offloading." }, "hardware": { "minimum_vram_gb_q4": 8, "comfortable_vram_gb_q4": 12, "recommended_gpu": "RTX 4060 8GB or better", "local_fit_note": "Good small-local-model candidate for 8GB+ GPUs at Q4." }, "routing_guidance": { "green": "Prefer local inference first if the selected GPU has comfortable headroom.", "yellow": "Reduce context, use lower quantization, try a smaller model, or expect RAM/offload slowdown.", "red": "Use a smaller local model, larger GPU, cloud GPU fallback, or compare API/cloud costs.", "commercial_option_ids": [], "ranking_influenced_by_affiliate": false }, "recommended_applications": [ { "id": "hermes-kellerrechner-agent", "name": "Hermes / Kellerrechner Agent", "category": "Personal AI assistant" }, { "id": "telegram-ai-bot", "name": "Telegram AI Bot", "category": "Chat and automation bot" }, { "id": "knowledge-vault-agent", "name": "Obsidian / Knowledge Vault Agent", "category": "Knowledge and memory agent" }, { "id": "social-publishing-approval-worker", "name": "Social Publishing Approval Worker", "category": "Marketing automation agent" } ], "sources": [ { "label": "Ollama library", "url": "https://ollama.com/library/qwen3" } ], "last_verified": "2026-05-19", "confidence": "medium" }, { "id": "qwen3-coder-30b", "model_name": "Qwen3-Coder 30B-A3B", "family": "Qwen3-Coder", "provider": "Alibaba", "license_type": "open-weight", "parameters_billions": 30, "architecture": "moe", "ollama_tag": "qwen3-coder:30b", "recommended_quantization": "Q4_K_M", "best_for": [ "chat", "coding", "agents/tool workflows", "reasoning", "Agentic coding, repository-scale local code review, and tool-heavy development loops" ], "weak_for": [ "vision/image understanding", "Long context leaves little headroom on single 24GB GPUs" ], "not_for": [ "image understanding / multimodal vision" ], "agent_readiness_score": 94, "context": { "minimum_context_tokens": 32768, "typical_context_tokens": 256000, "context_fit_note": "Large context-capable on paper; verify runtime memory and backend limits." }, "memory_estimates_gb": { "quantization": { "q4": 19, "q5": 23.75, "q8": 38, "fp16": 76 }, "q4_context_profiles": { "quick": 17.1, "coding": 19, "repo": 21.85, "rag": 25.65, "agent": 27.55, "extreme": 32.3 }, "caveat": "Estimates are practical planning estimates. Real use varies by backend, KV cache, context length, drivers, quantization file, and offloading." }, "hardware": { "minimum_vram_gb_q4": 24, "comfortable_vram_gb_q4": 32, "recommended_gpu": "RTX 3090/4090 24GB minimum, 32GB+ preferred", "local_fit_note": "Workstation-local candidate. Prefer 32GB+ VRAM for agents or long context." }, "routing_guidance": { "green": "Prefer local inference first if the selected GPU has comfortable headroom.", "yellow": "Reduce context, use lower quantization, try a smaller model, or expect RAM/offload slowdown.", "red": "Use a smaller local model, larger GPU, cloud GPU fallback, or compare API/cloud costs.", "commercial_option_ids": [ "apiroute-cloud-api-comparison" ], "ranking_influenced_by_affiliate": false }, "recommended_applications": [], "sources": [ { "label": "Ollama library", "url": "https://ollama.com/library/qwen3-coder" }, { "label": "Qwen blog", "url": "https://qwenlm.github.io/blog/qwen3-coder/" } ], "last_verified": "2026-05-26", "confidence": "high" }, { "id": "qwen3-5-9b", "model_name": "Qwen3.5 9B", "family": "Qwen3.5", "provider": "Alibaba", "license_type": "open-weight", "parameters_billions": 9, "architecture": "dense", "ollama_tag": "qwen3.5:9b", "recommended_quantization": "Q4_K_M", "best_for": [ "chat", "coding", "agents/tool workflows", "vision/image understanding", "reasoning", "Modern multimodal local assistant, agent experiments, and coding support on mainstream GPUs" ], "weak_for": [ "Still a small model for large repo-scale coding tasks" ], "not_for": [], "agent_readiness_score": 88, "context": { "minimum_context_tokens": 32768, "typical_context_tokens": 256000, "context_fit_note": "Large context-capable on paper; verify runtime memory and backend limits." }, "memory_estimates_gb": { "quantization": { "q4": 6.6, "q5": 8.25, "q8": 13.2, "fp16": 26.4 }, "q4_context_profiles": { "quick": 5.94, "coding": 6.6, "repo": 7.59, "rag": 8.91, "agent": 9.57, "extreme": 11.22 }, "caveat": "Estimates are practical planning estimates. Real use varies by backend, KV cache, context length, drivers, quantization file, and offloading." }, "hardware": { "minimum_vram_gb_q4": 8, "comfortable_vram_gb_q4": 12, "recommended_gpu": "RTX 4060 8GB minimum, RTX 3060 12GB preferred", "local_fit_note": "Good small-local-model candidate for 8GB+ GPUs at Q4." }, "routing_guidance": { "green": "Prefer local inference first if the selected GPU has comfortable headroom.", "yellow": "Reduce context, use lower quantization, try a smaller model, or expect RAM/offload slowdown.", "red": "Use a smaller local model, larger GPU, cloud GPU fallback, or compare API/cloud costs.", "commercial_option_ids": [], "ranking_influenced_by_affiliate": false }, "recommended_applications": [], "sources": [ { "label": "Ollama library", "url": "https://ollama.com/library/qwen3.5" }, { "label": "Qwen research index", "url": "https://qwen.ai/research" } ], "last_verified": "2026-05-26", "confidence": "medium" }, { "id": "qwen3-5-27b", "model_name": "Qwen3.5 27B", "family": "Qwen3.5", "provider": "Alibaba", "license_type": "open-weight", "parameters_billions": 27, "architecture": "dense", "ollama_tag": "qwen3.5:27b", "recommended_quantization": "Q4_K_M", "best_for": [ "chat", "coding", "agents/tool workflows", "vision/image understanding", "reasoning", "24GB-class multimodal agent, coding assistant, and reasoning workloads" ], "weak_for": [ "Long multimodal context can exceed single 24GB headroom" ], "not_for": [], "agent_readiness_score": 92, "context": { "minimum_context_tokens": 32768, "typical_context_tokens": 256000, "context_fit_note": "Large context-capable on paper; verify runtime memory and backend limits." }, "memory_estimates_gb": { "quantization": { "q4": 17, "q5": 21.25, "q8": 34, "fp16": 68 }, "q4_context_profiles": { "quick": 15.3, "coding": 17, "repo": 19.55, "rag": 22.95, "agent": 24.65, "extreme": 28.9 }, "caveat": "Estimates are practical planning estimates. Real use varies by backend, KV cache, context length, drivers, quantization file, and offloading." }, "hardware": { "minimum_vram_gb_q4": 24, "comfortable_vram_gb_q4": 32, "recommended_gpu": "RTX 3090/4090 24GB minimum, 32GB+ preferred", "local_fit_note": "Workstation-local candidate. Prefer 32GB+ VRAM for agents or long context." }, "routing_guidance": { "green": "Prefer local inference first if the selected GPU has comfortable headroom.", "yellow": "Reduce context, use lower quantization, try a smaller model, or expect RAM/offload slowdown.", "red": "Use a smaller local model, larger GPU, cloud GPU fallback, or compare API/cloud costs.", "commercial_option_ids": [ "apiroute-cloud-api-comparison" ], "ranking_influenced_by_affiliate": false }, "recommended_applications": [], "sources": [ { "label": "Ollama library", "url": "https://ollama.com/library/qwen3.5" }, { "label": "Qwen research index", "url": "https://qwen.ai/research" } ], "last_verified": "2026-05-26", "confidence": "medium" }, { "id": "deepseek-r1-distill-qwen-8b", "model_name": "DeepSeek R1 Distill Qwen 8B", "family": "DeepSeek R1 Distill", "provider": "DeepSeek", "license_type": "open-weight", "parameters_billions": 8, "architecture": "dense", "ollama_tag": "deepseek-r1:8b", "recommended_quantization": "Q4_K_M", "best_for": [ "chat", "coding", "reasoning", "Local reasoning experiments and step-by-step technical analysis" ], "weak_for": [ "agents/tool workflows", "vision/image understanding", "Verbose reasoning can slow simple agent workflows" ], "not_for": [ "image understanding / multimodal vision", "long-running tool agents" ], "agent_readiness_score": 72, "context": { "minimum_context_tokens": 8192, "typical_context_tokens": 128000, "context_fit_note": "Large context-capable on paper; verify runtime memory and backend limits." }, "memory_estimates_gb": { "quantization": { "q4": 6, "q5": 7.5, "q8": 12, "fp16": 24 }, "q4_context_profiles": { "quick": 5.4, "coding": 6, "repo": 6.9, "rag": 8.1, "agent": 8.7, "extreme": 10.2 }, "caveat": "Estimates are practical planning estimates. Real use varies by backend, KV cache, context length, drivers, quantization file, and offloading." }, "hardware": { "minimum_vram_gb_q4": 8, "comfortable_vram_gb_q4": 12, "recommended_gpu": "RTX 4060 8GB or better", "local_fit_note": "Good small-local-model candidate for 8GB+ GPUs at Q4." }, "routing_guidance": { "green": "Prefer local inference first if the selected GPU has comfortable headroom.", "yellow": "Reduce context, use lower quantization, try a smaller model, or expect RAM/offload slowdown.", "red": "Use a smaller local model, larger GPU, cloud GPU fallback, or compare API/cloud costs.", "commercial_option_ids": [], "ranking_influenced_by_affiliate": false }, "recommended_applications": [], "sources": [ { "label": "Ollama library", "url": "https://ollama.com/library/deepseek-r1" } ], "last_verified": "2026-05-19", "confidence": "high" }, { "id": "deepseek-r1-distill-qwen-14b", "model_name": "DeepSeek R1 Distill Qwen 14B", "family": "DeepSeek R1 Distill", "provider": "DeepSeek", "license_type": "open-weight", "parameters_billions": 14, "architecture": "dense", "ollama_tag": "deepseek-r1:14b", "recommended_quantization": "Q4_K_M", "best_for": [ "chat", "coding", "reasoning", "Local reasoning and debugging on 12GB/16GB GPUs" ], "weak_for": [ "agents/tool workflows", "vision/image understanding", "Less ergonomic for fast Telegram-style assistant responses" ], "not_for": [ "image understanding / multimodal vision", "long-running tool agents" ], "agent_readiness_score": 76, "context": { "minimum_context_tokens": 8192, "typical_context_tokens": 128000, "context_fit_note": "Large context-capable on paper; verify runtime memory and backend limits." }, "memory_estimates_gb": { "quantization": { "q4": 10.5, "q5": 13.13, "q8": 21, "fp16": 42 }, "q4_context_profiles": { "quick": 9.45, "coding": 10.5, "repo": 12.07, "rag": 14.18, "agent": 15.22, "extreme": 17.85 }, "caveat": "Estimates are practical planning estimates. Real use varies by backend, KV cache, context length, drivers, quantization file, and offloading." }, "hardware": { "minimum_vram_gb_q4": 12, "comfortable_vram_gb_q4": 16, "recommended_gpu": "RTX 3060 12GB minimum, 16GB+ preferred", "local_fit_note": "Practical 12GB local-agent candidate at Q4 with headroom checks." }, "routing_guidance": { "green": "Prefer local inference first if the selected GPU has comfortable headroom.", "yellow": "Reduce context, use lower quantization, try a smaller model, or expect RAM/offload slowdown.", "red": "Use a smaller local model, larger GPU, cloud GPU fallback, or compare API/cloud costs.", "commercial_option_ids": [], "ranking_influenced_by_affiliate": false }, "recommended_applications": [], "sources": [ { "label": "Ollama library", "url": "https://ollama.com/library/deepseek-r1" } ], "last_verified": "2026-05-19", "confidence": "high" }, { "id": "deepseek-r1-distill-qwen-32b", "model_name": "DeepSeek R1 Distill Qwen 32B", "family": "DeepSeek R1 Distill", "provider": "DeepSeek", "license_type": "open-weight", "parameters_billions": 32, "architecture": "dense", "ollama_tag": "deepseek-r1:32b", "recommended_quantization": "Q4_K_M", "best_for": [ "chat", "coding", "reasoning", "Heavy local reasoning on 24GB GPUs" ], "weak_for": [ "agents/tool workflows", "vision/image understanding", "Tight VRAM headroom and slower agent loops" ], "not_for": [ "image understanding / multimodal vision", "long-running tool agents" ], "agent_readiness_score": 78, "context": { "minimum_context_tokens": 8192, "typical_context_tokens": 128000, "context_fit_note": "Large context-capable on paper; verify runtime memory and backend limits." }, "memory_estimates_gb": { "quantization": { "q4": 21, "q5": 26.25, "q8": 42, "fp16": 84 }, "q4_context_profiles": { "quick": 18.9, "coding": 21, "repo": 24.15, "rag": 28.35, "agent": 30.45, "extreme": 35.7 }, "caveat": "Estimates are practical planning estimates. Real use varies by backend, KV cache, context length, drivers, quantization file, and offloading." }, "hardware": { "minimum_vram_gb_q4": 24, "comfortable_vram_gb_q4": 32, "recommended_gpu": "RTX 3090/4090 24GB minimum, 32GB+ preferred", "local_fit_note": "Workstation-local candidate. Prefer 32GB+ VRAM for agents or long context." }, "routing_guidance": { "green": "Prefer local inference first if the selected GPU has comfortable headroom.", "yellow": "Reduce context, use lower quantization, try a smaller model, or expect RAM/offload slowdown.", "red": "Use a smaller local model, larger GPU, cloud GPU fallback, or compare API/cloud costs.", "commercial_option_ids": [ "apiroute-cloud-api-comparison" ], "ranking_influenced_by_affiliate": false }, "recommended_applications": [ { "id": "desktop-multi-agent-command-center", "name": "Desktop Multi-Agent Command Center", "category": "Agent orchestration" } ], "sources": [ { "label": "Ollama library", "url": "https://ollama.com/library/deepseek-r1" } ], "last_verified": "2026-05-19", "confidence": "medium" }, { "id": "gemma-3-4b", "model_name": "Gemma 3 4B", "family": "Gemma", "provider": "Google", "license_type": "open-weight", "parameters_billions": 4, "architecture": "dense", "ollama_tag": "gemma3:4b", "recommended_quantization": "Q4_K_M", "best_for": [ "chat", "agents/tool workflows", "vision/image understanding", "Small multimodal local assistant and low-resource setups" ], "weak_for": [ "coding", "reasoning", "Limited quality for coding and complex tasks" ], "not_for": [ "coding-heavy workloads", "deep multi-step reasoning" ], "agent_readiness_score": 70, "context": { "minimum_context_tokens": 8192, "typical_context_tokens": 128000, "context_fit_note": "Large context-capable on paper; verify runtime memory and backend limits." }, "memory_estimates_gb": { "quantization": { "q4": 3.5, "q5": 4.38, "q8": 7, "fp16": 14 }, "q4_context_profiles": { "quick": 3.15, "coding": 3.5, "repo": 4.02, "rag": 4.73, "agent": 5.08, "extreme": 5.95 }, "caveat": "Estimates are practical planning estimates. Real use varies by backend, KV cache, context length, drivers, quantization file, and offloading." }, "hardware": { "minimum_vram_gb_q4": 6, "comfortable_vram_gb_q4": 8, "recommended_gpu": "6GB+ VRAM GPU", "local_fit_note": "Good small-local-model candidate for 8GB+ GPUs at Q4." }, "routing_guidance": { "green": "Prefer local inference first if the selected GPU has comfortable headroom.", "yellow": "Reduce context, use lower quantization, try a smaller model, or expect RAM/offload slowdown.", "red": "Use a smaller local model, larger GPU, cloud GPU fallback, or compare API/cloud costs.", "commercial_option_ids": [], "ranking_influenced_by_affiliate": false }, "recommended_applications": [], "sources": [ { "label": "Ollama library", "url": "https://ollama.com/library/gemma3" } ], "last_verified": "2026-05-19", "confidence": "high" }, { "id": "gemma-3-12b", "model_name": "Gemma 3 12B", "family": "Gemma", "provider": "Google", "license_type": "open-weight", "parameters_billions": 12, "architecture": "dense", "ollama_tag": "gemma3:12b", "recommended_quantization": "Q4_K_M", "best_for": [ "chat", "agents/tool workflows", "vision/image understanding", "reasoning", "Balanced multimodal local chat on 12GB+ GPUs" ], "weak_for": [ "coding", "Not primarily a coding model" ], "not_for": [ "coding-heavy workloads" ], "agent_readiness_score": 78, "context": { "minimum_context_tokens": 8192, "typical_context_tokens": 128000, "context_fit_note": "Large context-capable on paper; verify runtime memory and backend limits." }, "memory_estimates_gb": { "quantization": { "q4": 9, "q5": 11.25, "q8": 18, "fp16": 36 }, "q4_context_profiles": { "quick": 8.1, "coding": 9, "repo": 10.35, "rag": 12.15, "agent": 13.05, "extreme": 15.3 }, "caveat": "Estimates are practical planning estimates. Real use varies by backend, KV cache, context length, drivers, quantization file, and offloading." }, "hardware": { "minimum_vram_gb_q4": 12, "comfortable_vram_gb_q4": 16, "recommended_gpu": "RTX 3060 12GB or better", "local_fit_note": "Practical 12GB local-agent candidate at Q4 with headroom checks." }, "routing_guidance": { "green": "Prefer local inference first if the selected GPU has comfortable headroom.", "yellow": "Reduce context, use lower quantization, try a smaller model, or expect RAM/offload slowdown.", "red": "Use a smaller local model, larger GPU, cloud GPU fallback, or compare API/cloud costs.", "commercial_option_ids": [], "ranking_influenced_by_affiliate": false }, "recommended_applications": [], "sources": [ { "label": "Ollama library", "url": "https://ollama.com/library/gemma3" } ], "last_verified": "2026-05-19", "confidence": "high" }, { "id": "gemma-3-27b", "model_name": "Gemma 3 27B", "family": "Gemma", "provider": "Google", "license_type": "open-weight", "parameters_billions": 27, "architecture": "dense", "ollama_tag": "gemma3:27b", "recommended_quantization": "Q4_K_M", "best_for": [ "chat", "agents/tool workflows", "vision/image understanding", "reasoning", "High-quality multimodal local assistant on 24GB GPUs" ], "weak_for": [ "coding", "Less specialized for code than Qwen Coder" ], "not_for": [ "coding-heavy workloads" ], "agent_readiness_score": 84, "context": { "minimum_context_tokens": 8192, "typical_context_tokens": 128000, "context_fit_note": "Large context-capable on paper; verify runtime memory and backend limits." }, "memory_estimates_gb": { "quantization": { "q4": 18, "q5": 22.5, "q8": 36, "fp16": 72 }, "q4_context_profiles": { "quick": 16.2, "coding": 18, "repo": 20.7, "rag": 24.3, "agent": 26.1, "extreme": 30.6 }, "caveat": "Estimates are practical planning estimates. Real use varies by backend, KV cache, context length, drivers, quantization file, and offloading." }, "hardware": { "minimum_vram_gb_q4": 24, "comfortable_vram_gb_q4": 32, "recommended_gpu": "RTX 3090/4090 24GB or better", "local_fit_note": "Workstation-local candidate. Prefer 32GB+ VRAM for agents or long context." }, "routing_guidance": { "green": "Prefer local inference first if the selected GPU has comfortable headroom.", "yellow": "Reduce context, use lower quantization, try a smaller model, or expect RAM/offload slowdown.", "red": "Use a smaller local model, larger GPU, cloud GPU fallback, or compare API/cloud costs.", "commercial_option_ids": [ "apiroute-cloud-api-comparison" ], "ranking_influenced_by_affiliate": false }, "recommended_applications": [], "sources": [ { "label": "Ollama library", "url": "https://ollama.com/library/gemma3" } ], "last_verified": "2026-05-19", "confidence": "high" }, { "id": "gemma-4-e4b", "model_name": "Gemma 4 E4B", "family": "Gemma", "provider": "Google", "license_type": "open-weight", "parameters_billions": 4, "architecture": "moe", "ollama_tag": "gemma4:e4b", "recommended_quantization": "Q4_K_M", "best_for": [ "chat", "coding", "agents/tool workflows", "vision/image understanding", "reasoning", "Efficient multimodal local assistant and edge-style agent workflows" ], "weak_for": [ "Smaller effective model; not ideal for deep repository-scale coding" ], "not_for": [], "agent_readiness_score": 83, "context": { "minimum_context_tokens": 32768, "typical_context_tokens": 128000, "context_fit_note": "Large context-capable on paper; verify runtime memory and backend limits." }, "memory_estimates_gb": { "quantization": { "q4": 9.6, "q5": 12, "q8": 19.2, "fp16": 38.4 }, "q4_context_profiles": { "quick": 8.64, "coding": 9.6, "repo": 11.04, "rag": 12.96, "agent": 13.92, "extreme": 16.32 }, "caveat": "Estimates are practical planning estimates. Real use varies by backend, KV cache, context length, drivers, quantization file, and offloading." }, "hardware": { "minimum_vram_gb_q4": 12, "comfortable_vram_gb_q4": 16, "recommended_gpu": "RTX 3060 12GB minimum, 16GB+ preferred", "local_fit_note": "Practical 12GB local-agent candidate at Q4 with headroom checks." }, "routing_guidance": { "green": "Prefer local inference first if the selected GPU has comfortable headroom.", "yellow": "Reduce context, use lower quantization, try a smaller model, or expect RAM/offload slowdown.", "red": "Use a smaller local model, larger GPU, cloud GPU fallback, or compare API/cloud costs.", "commercial_option_ids": [], "ranking_influenced_by_affiliate": false }, "recommended_applications": [], "sources": [ { "label": "Ollama library", "url": "https://ollama.com/library/gemma4" }, { "label": "Google DeepMind Gemma 4", "url": "https://deepmind.google/models/gemma/gemma-4/" } ], "last_verified": "2026-05-26", "confidence": "high" }, { "id": "gemma-4-31b", "model_name": "Gemma 4 31B", "family": "Gemma", "provider": "Google", "license_type": "open-weight", "parameters_billions": 31, "architecture": "dense", "ollama_tag": "gemma4:31b", "recommended_quantization": "Q4_K_M", "best_for": [ "chat", "coding", "agents/tool workflows", "vision/image understanding", "reasoning", "High-quality multimodal reasoning, coding assistants, and local-first agent workflows" ], "weak_for": [ "Single 24GB GPUs have limited headroom for long context" ], "not_for": [], "agent_readiness_score": 91, "context": { "minimum_context_tokens": 32768, "typical_context_tokens": 256000, "context_fit_note": "Large context-capable on paper; verify runtime memory and backend limits." }, "memory_estimates_gb": { "quantization": { "q4": 20, "q5": 25, "q8": 40, "fp16": 80 }, "q4_context_profiles": { "quick": 18, "coding": 20, "repo": 23, "rag": 27, "agent": 29, "extreme": 34 }, "caveat": "Estimates are practical planning estimates. Real use varies by backend, KV cache, context length, drivers, quantization file, and offloading." }, "hardware": { "minimum_vram_gb_q4": 24, "comfortable_vram_gb_q4": 32, "recommended_gpu": "RTX 3090/4090 24GB minimum, 32GB+ preferred", "local_fit_note": "Workstation-local candidate. Prefer 32GB+ VRAM for agents or long context." }, "routing_guidance": { "green": "Prefer local inference first if the selected GPU has comfortable headroom.", "yellow": "Reduce context, use lower quantization, try a smaller model, or expect RAM/offload slowdown.", "red": "Use a smaller local model, larger GPU, cloud GPU fallback, or compare API/cloud costs.", "commercial_option_ids": [ "apiroute-cloud-api-comparison" ], "ranking_influenced_by_affiliate": false }, "recommended_applications": [], "sources": [ { "label": "Ollama library", "url": "https://ollama.com/library/gemma4" }, { "label": "Google DeepMind Gemma 4", "url": "https://deepmind.google/models/gemma/gemma-4/" } ], "last_verified": "2026-05-26", "confidence": "high" }, { "id": "mistral-7b", "model_name": "Mistral 7B", "family": "Mistral", "provider": "Mistral AI", "license_type": "open-weight", "parameters_billions": 7, "architecture": "dense", "ollama_tag": "mistral:7b", "recommended_quantization": "Q4_K_M", "best_for": [ "chat", "agents/tool workflows", "Fast local chat and simple agent tasks" ], "weak_for": [ "coding", "vision/image understanding", "reasoning", "Older/smaller than newer Qwen/Gemma alternatives" ], "not_for": [ "image understanding / multimodal vision", "coding-heavy workloads", "deep multi-step reasoning" ], "agent_readiness_score": 74, "context": { "minimum_context_tokens": 8192, "typical_context_tokens": 32768, "context_fit_note": "Good practical context range for coding, chat, and moderate RAG." }, "memory_estimates_gb": { "quantization": { "q4": 5.5, "q5": 6.88, "q8": 11, "fp16": 22 }, "q4_context_profiles": { "quick": 4.95, "coding": 5.5, "repo": 6.32, "rag": 7.43, "agent": 7.97, "extreme": 9.35 }, "caveat": "Estimates are practical planning estimates. Real use varies by backend, KV cache, context length, drivers, quantization file, and offloading." }, "hardware": { "minimum_vram_gb_q4": 8, "comfortable_vram_gb_q4": 12, "recommended_gpu": "RTX 4060 8GB or better", "local_fit_note": "Good small-local-model candidate for 8GB+ GPUs at Q4." }, "routing_guidance": { "green": "Prefer local inference first if the selected GPU has comfortable headroom.", "yellow": "Reduce context, use lower quantization, try a smaller model, or expect RAM/offload slowdown.", "red": "Use a smaller local model, larger GPU, cloud GPU fallback, or compare API/cloud costs.", "commercial_option_ids": [], "ranking_influenced_by_affiliate": false }, "recommended_applications": [ { "id": "telegram-ai-bot", "name": "Telegram AI Bot", "category": "Chat and automation bot" } ], "sources": [ { "label": "Ollama library", "url": "https://ollama.com/library/mistral" } ], "last_verified": "2026-05-19", "confidence": "high" }, { "id": "devstral-small-2-24b", "model_name": "Devstral Small 2 24B", "family": "Devstral", "provider": "Mistral AI", "license_type": "open-weight", "parameters_billions": 24, "architecture": "dense", "ollama_tag": "devstral-small-2", "recommended_quantization": "Q4_K_M", "best_for": [ "chat", "coding", "agents/tool workflows", "vision/image understanding", "reasoning", "Software engineering agents, repo navigation, patch planning, and local coding workflows" ], "weak_for": [ "Large-context coding work is tight below 24GB VRAM" ], "not_for": [], "agent_readiness_score": 91, "context": { "minimum_context_tokens": 32768, "typical_context_tokens": 384000, "context_fit_note": "Large context-capable on paper; verify runtime memory and backend limits." }, "memory_estimates_gb": { "quantization": { "q4": 15, "q5": 18.75, "q8": 30, "fp16": 60 }, "q4_context_profiles": { "quick": 13.5, "coding": 15, "repo": 17.25, "rag": 20.25, "agent": 21.75, "extreme": 25.5 }, "caveat": "Estimates are practical planning estimates. Real use varies by backend, KV cache, context length, drivers, quantization file, and offloading." }, "hardware": { "minimum_vram_gb_q4": 16, "comfortable_vram_gb_q4": 24, "recommended_gpu": "16GB VRAM minimum, 24GB+ preferred", "local_fit_note": "Workstation-local candidate. Prefer 24GB+ VRAM for agents or long context." }, "routing_guidance": { "green": "Prefer local inference first if the selected GPU has comfortable headroom.", "yellow": "Reduce context, use lower quantization, try a smaller model, or expect RAM/offload slowdown.", "red": "Use a smaller local model, larger GPU, cloud GPU fallback, or compare API/cloud costs.", "commercial_option_ids": [ "apiroute-cloud-api-comparison" ], "ranking_influenced_by_affiliate": false }, "recommended_applications": [], "sources": [ { "label": "Ollama library", "url": "https://ollama.com/library/devstral-small-2" }, { "label": "Mistral AI model card", "url": "https://docs.mistral.ai/models/model-cards/devstral-small-2-25-12" }, { "label": "Mistral AI announcement", "url": "https://mistral.ai/news/devstral-2-vibe-cli" } ], "last_verified": "2026-05-26", "confidence": "high" }, { "id": "mixtral-8x7b", "model_name": "Mixtral 8x7B", "family": "Mixtral", "provider": "Mistral AI", "license_type": "open-weight", "parameters_billions": 46.7, "architecture": "moe", "ollama_tag": "mixtral:8x7b", "recommended_quantization": "Q4_K_M", "best_for": [ "chat", "coding", "agents/tool workflows", "reasoning", "MoE local reasoning/chat when enough VRAM is available" ], "weak_for": [ "vision/image understanding", "Not practical for 24GB single-GPU setups without offload" ], "not_for": [ "image understanding / multimodal vision" ], "agent_readiness_score": 82, "context": { "minimum_context_tokens": 8192, "typical_context_tokens": 32768, "context_fit_note": "Good practical context range for coding, chat, and moderate RAG." }, "memory_estimates_gb": { "quantization": { "q4": 28, "q5": 35, "q8": 56, "fp16": 112 }, "q4_context_profiles": { "quick": 25.2, "coding": 28, "repo": 32.2, "rag": 37.8, "agent": 40.6, "extreme": 47.6 }, "caveat": "Estimates are practical planning estimates. Real use varies by backend, KV cache, context length, drivers, quantization file, and offloading." }, "hardware": { "minimum_vram_gb_q4": 32, "comfortable_vram_gb_q4": 48, "recommended_gpu": "32GB+ VRAM or CPU/RAM offload", "local_fit_note": "Large local model. Prefer 48GB+ VRAM, multi-GPU, cloud GPU, or hosted API fallback." }, "routing_guidance": { "green": "Prefer local inference first if the selected GPU has comfortable headroom.", "yellow": "Reduce context, use lower quantization, try a smaller model, or expect RAM/offload slowdown.", "red": "Use a smaller local model, larger GPU, cloud GPU fallback, or compare API/cloud costs.", "commercial_option_ids": [ "runpod-cloud-gpu-fallback", "apiroute-cloud-api-comparison" ], "ranking_influenced_by_affiliate": false }, "recommended_applications": [], "sources": [ { "label": "Ollama library", "url": "https://ollama.com/library/mixtral" } ], "last_verified": "2026-05-19", "confidence": "high" }, { "id": "phi-4-14b", "model_name": "Phi-4 14B", "family": "Phi", "provider": "Microsoft", "license_type": "open-weight", "parameters_billions": 14, "architecture": "dense", "ollama_tag": "phi4:14b", "recommended_quantization": "Q4_K_M", "best_for": [ "chat", "coding", "agents/tool workflows", "reasoning", "Compact reasoning and technical assistant on 12GB/16GB GPUs" ], "weak_for": [ "vision/image understanding", "Smaller ecosystem than Llama/Qwen families" ], "not_for": [ "image understanding / multimodal vision" ], "agent_readiness_score": 82, "context": { "minimum_context_tokens": 8192, "typical_context_tokens": 16384, "context_fit_note": "Short-to-moderate context; avoid large document or whole-repo workflows." }, "memory_estimates_gb": { "quantization": { "q4": 10.5, "q5": 13.13, "q8": 21, "fp16": 42 }, "q4_context_profiles": { "quick": 9.45, "coding": 10.5, "repo": 12.07, "rag": 14.18, "agent": 15.22, "extreme": 17.85 }, "caveat": "Estimates are practical planning estimates. Real use varies by backend, KV cache, context length, drivers, quantization file, and offloading." }, "hardware": { "minimum_vram_gb_q4": 12, "comfortable_vram_gb_q4": 16, "recommended_gpu": "RTX 3060 12GB minimum, 16GB+ preferred", "local_fit_note": "Practical 12GB local-agent candidate at Q4 with headroom checks." }, "routing_guidance": { "green": "Prefer local inference first if the selected GPU has comfortable headroom.", "yellow": "Reduce context, use lower quantization, try a smaller model, or expect RAM/offload slowdown.", "red": "Use a smaller local model, larger GPU, cloud GPU fallback, or compare API/cloud costs.", "commercial_option_ids": [], "ranking_influenced_by_affiliate": false }, "recommended_applications": [ { "id": "local-coding-agent", "name": "Local Coding Agent", "category": "Coding agent" } ], "sources": [ { "label": "Ollama library", "url": "https://ollama.com/library/phi4" } ], "last_verified": "2026-05-19", "confidence": "high" }, { "id": "gpt-oss-20b", "model_name": "gpt-oss 20B", "family": "gpt-oss", "provider": "OpenAI", "license_type": "open-weight", "parameters_billions": 20, "architecture": "moe", "ollama_tag": "gpt-oss:20b", "recommended_quantization": "MXFP4 native", "best_for": [ "chat", "coding", "agents/tool workflows", "reasoning", "Local reasoning, agent planning, and tool-use workflows on 16GB+ GPUs" ], "weak_for": [ "vision/image understanding", "12GB GPUs need offload or smaller fallback models" ], "not_for": [ "image understanding / multimodal vision" ], "agent_readiness_score": 89, "context": { "minimum_context_tokens": 32768, "typical_context_tokens": 128000, "context_fit_note": "Large context-capable on paper; verify runtime memory and backend limits." }, "memory_estimates_gb": { "quantization": { "q4": 14, "q5": 17.5, "q8": 28, "fp16": 56 }, "q4_context_profiles": { "quick": 12.6, "coding": 14, "repo": 16.1, "rag": 18.9, "agent": 20.3, "extreme": 23.8 }, "caveat": "Estimates are practical planning estimates. Real use varies by backend, KV cache, context length, drivers, quantization file, and offloading." }, "hardware": { "minimum_vram_gb_q4": 16, "comfortable_vram_gb_q4": 24, "recommended_gpu": "16GB VRAM minimum, 24GB+ preferred", "local_fit_note": "Workstation-local candidate. Prefer 24GB+ VRAM for agents or long context." }, "routing_guidance": { "green": "Prefer local inference first if the selected GPU has comfortable headroom.", "yellow": "Reduce context, use lower quantization, try a smaller model, or expect RAM/offload slowdown.", "red": "Use a smaller local model, larger GPU, cloud GPU fallback, or compare API/cloud costs.", "commercial_option_ids": [ "apiroute-cloud-api-comparison" ], "ranking_influenced_by_affiliate": false }, "recommended_applications": [], "sources": [ { "label": "Ollama library", "url": "https://ollama.com/library/gpt-oss" }, { "label": "OpenAI gpt-oss", "url": "https://cdn.openai.com/pdf/419b6906-9da6-406c-a19d-1bb078ac7637/oai_gpt-oss_model_card.pdf" } ], "last_verified": "2026-05-26", "confidence": "high" }, { "id": "gpt-oss-120b", "model_name": "gpt-oss 120B", "family": "gpt-oss", "provider": "OpenAI", "license_type": "open-weight", "parameters_billions": 120, "architecture": "moe", "ollama_tag": "gpt-oss:120b", "recommended_quantization": "MXFP4 native", "best_for": [ "chat", "coding", "agents/tool workflows", "reasoning", "Large local reasoning servers, heavy agent orchestration, and high-end homelab inference" ], "weak_for": [ "vision/image understanding", "Not realistic for consumer single-GPU setups below 80GB-class memory" ], "not_for": [ "image understanding / multimodal vision" ], "agent_readiness_score": 93, "context": { "minimum_context_tokens": 32768, "typical_context_tokens": 128000, "context_fit_note": "Large context-capable on paper; verify runtime memory and backend limits." }, "memory_estimates_gb": { "quantization": { "q4": 65, "q5": 81.25, "q8": 130, "fp16": 260 }, "q4_context_profiles": { "quick": 58.5, "coding": 65, "repo": 74.75, "rag": 87.75, "agent": 94.25, "extreme": 110.5 }, "caveat": "Estimates are practical planning estimates. Real use varies by backend, KV cache, context length, drivers, quantization file, and offloading." }, "hardware": { "minimum_vram_gb_q4": 80, "comfortable_vram_gb_q4": 96, "recommended_gpu": "80GB+ VRAM server GPU or multi-GPU setup", "local_fit_note": "Large local model. Prefer 48GB+ VRAM, multi-GPU, cloud GPU, or hosted API fallback." }, "routing_guidance": { "green": "Prefer local inference first if the selected GPU has comfortable headroom.", "yellow": "Reduce context, use lower quantization, try a smaller model, or expect RAM/offload slowdown.", "red": "Use a smaller local model, larger GPU, cloud GPU fallback, or compare API/cloud costs.", "commercial_option_ids": [ "runpod-cloud-gpu-fallback", "apiroute-cloud-api-comparison" ], "ranking_influenced_by_affiliate": false }, "recommended_applications": [], "sources": [ { "label": "Ollama library", "url": "https://ollama.com/library/gpt-oss" }, { "label": "OpenAI gpt-oss", "url": "https://cdn.openai.com/pdf/419b6906-9da6-406c-a19d-1bb078ac7637/oai_gpt-oss_model_card.pdf" } ], "last_verified": "2026-05-26", "confidence": "high" } ] }