{"id":2199,"date":"2026-08-14T06:18:40","date_gmt":"2026-08-14T06:18:40","guid":{"rendered":"https:\/\/convly.ai\/?p=2199"},"modified":"2026-08-14T06:18:40","modified_gmt":"2026-08-14T06:18:40","slug":"ollama-gpt-oss-guide","status":"publish","type":"post","link":"https:\/\/convly.ai\/es\/ollama-gpt-oss-guide\/","title":{"rendered":"Ollama GPT OSS: Running Open-Source LLMs Locally (Guide)"},"content":{"rendered":"<div class=\"convly-tldr\"><strong>En resumen<\/strong><\/p>\n<ul>\n<li>OpenAI has not released a model family called &#8220;GPT-OSS&#8221;. The search term likely refers to running open-source alternatives (Llama 3, Mistral, Qwen) through Ollama.<\/li>\n<li>Ollama runs hundreds of open-weight models locally. Popular options: <code>ollama pull llama3.1:8b<\/code>, <code>ollama pull mistral:7b<\/code>, <code>ollama pull qwen2.5:7b<\/code>.<\/li>\n<li>VRAM needs: 7B models need 6-8 GB (Q4 quant), 13B models need 10-14 GB, 70B models need 40-48 GB. Use the <a href=\"https:\/\/convly.ai\/es\/llm-vram-calculator\/\">Calculadora de VRAM<\/a> to estimate.<\/li>\n<li>Quantized models (Q4, Q5) fit consumer GPUs with minimal quality loss. FP16 gives best quality but doubles VRAM.<\/li>\n<\/ul>\n<\/div>\n<p>If you searched &#8220;ollama gpt oss&#8221;, you likely want to run open-source large language models locally using Ollama\u2014but OpenAI has not released an open-weight model family called &#8220;GPT-OSS&#8221;. OpenAI released GPT-2 in 2019 as open-weight, but their recent models (GPT-4, GPT-4o, GPT-4.1) remain proprietary API-only products. What <em>no<\/em> exist: hundreds of open-weight models from Meta (Llama), Mistral AI, Alibaba (Qwen), and others that you can pull and run through Ollama on your own hardware. This guide covers which models work, the VRAM each size requires, how quantization affects quality, and how they compare.<\/p>\n<div id=\"ez-toc-container\" class=\"ez-toc-v2_0_86 counter-flat ez-toc-counter ez-toc-container-direction\">\n<label for=\"ez-toc-cssicon-toggle-item-6a7eeb54e9b8f\" class=\"ez-toc-cssicon-toggle-label\"><span class=\"\"><span class=\"eztoc-hide\" style=\"display:none;\">Alternar<\/span><span class=\"ez-toc-icon-toggle-span\"><svg style=\"fill: #000000;color:#000000\" xmlns=\"http:\/\/www.w3.org\/2000\/svg\" class=\"list-377408\" width=\"20px\" height=\"20px\" viewbox=\"0 0 24 24\" fill=\"none\"><path d=\"M6 6H4v2h2V6zm14 0H8v2h12V6zM4 11h2v2H4v-2zm16 0H8v2h12v-2zM4 16h2v2H4v-2zm16 0H8v2h12v-2z\" fill=\"currentColor\"><\/path><\/svg><svg style=\"fill: #000000;color:#000000\" class=\"arrow-unsorted-368013\" xmlns=\"http:\/\/www.w3.org\/2000\/svg\" width=\"10px\" height=\"10px\" viewbox=\"0 0 24 24\" version=\"1.2\" baseprofile=\"tiny\"><path d=\"M18.2 9.3l-6.2-6.3-6.2 6.3c-.2.2-.3.4-.3.7s.1.5.3.7c.2.2.4.3.7.3h11c.3 0 .5-.1.7-.3.2-.2.3-.5.3-.7s-.1-.5-.3-.7zM5.8 14.7l6.2 6.3 6.2-6.3c.2-.2.3-.5.3-.7s-.1-.5-.3-.7c-.2-.2-.4-.3-.7-.3h-11c-.3 0-.5.1-.7.3-.2.2-.3.5-.3.7s.1.5.3.7z\"\/><\/svg><\/span><\/span><\/label><input type=\"checkbox\"  id=\"ez-toc-cssicon-toggle-item-6a7eeb54e9b8f\"  aria-label=\"Alternar\" \/><nav><ul class='ez-toc-list ez-toc-list-level-1' ><li class='ez-toc-page-1'><a class=\"ez-toc-link ez-toc-heading-1\" href=\"https:\/\/convly.ai\/es\/ollama-gpt-oss-guide\/#What_Ollama_Actually_Runs\" >What Ollama Actually Runs<\/a><\/li><li class='ez-toc-page-1'><a class=\"ez-toc-link ez-toc-heading-2\" href=\"https:\/\/convly.ai\/es\/ollama-gpt-oss-guide\/#Open-Source_Model_Families_Available_in_Ollama\" >Open-Source Model Families Available in Ollama<\/a><\/li><li class='ez-toc-page-1'><a class=\"ez-toc-link ez-toc-heading-3\" href=\"https:\/\/convly.ai\/es\/ollama-gpt-oss-guide\/#Pulling_and_Running_a_Model\" >Pulling and Running a Model<\/a><\/li><li class='ez-toc-page-1'><a class=\"ez-toc-link ez-toc-heading-4\" href=\"https:\/\/convly.ai\/es\/ollama-gpt-oss-guide\/#Parameter_Sizes_and_VRAM_Requirements\" >Parameter Sizes and VRAM Requirements<\/a><\/li><li class='ez-toc-page-1'><a class=\"ez-toc-link ez-toc-heading-5\" href=\"https:\/\/convly.ai\/es\/ollama-gpt-oss-guide\/#Quantization_Quality_vs_VRAM_Trade-Off\" >Quantization: Quality vs VRAM Trade-Off<\/a><\/li><li class='ez-toc-page-1'><a class=\"ez-toc-link ez-toc-heading-6\" href=\"https:\/\/convly.ai\/es\/ollama-gpt-oss-guide\/#Performance_Comparison_7B_Class\" >Performance Comparison: 7B Class<\/a><\/li><li class='ez-toc-page-1'><a class=\"ez-toc-link ez-toc-heading-7\" href=\"https:\/\/convly.ai\/es\/ollama-gpt-oss-guide\/#Which_Model_to_Pick\" >Which Model to Pick<\/a><\/li><li class='ez-toc-page-1'><a class=\"ez-toc-link ez-toc-heading-8\" href=\"https:\/\/convly.ai\/es\/ollama-gpt-oss-guide\/#When_to_Self-Host_vs_Use_an_API\" >When to Self-Host vs Use an API<\/a><\/li><li class='ez-toc-page-1'><a class=\"ez-toc-link ez-toc-heading-9\" href=\"https:\/\/convly.ai\/es\/ollama-gpt-oss-guide\/#Platform-Specific_Notes\" >Platform-Specific Notes<\/a><\/li><li class='ez-toc-page-1'><a class=\"ez-toc-link ez-toc-heading-10\" href=\"https:\/\/convly.ai\/es\/ollama-gpt-oss-guide\/#Frequently_Asked_Questions\" >Preguntas frecuentes<\/a><\/li><\/ul><\/nav><\/div>\n<h2><span class=\"ez-toc-section\" id=\"What_Ollama_Actually_Runs\"><\/span>What Ollama Actually Runs<span class=\"ez-toc-section-end\"><\/span><\/h2>\n<p>Ollama is a local inference engine that downloads, quantizes, and runs open-weight LLMs on macOS, Linux, and Windows. It does not host OpenAI models. The <a href=\"https:\/\/convly.ai\/es\/ollama-models-list-2026\/\">Lista de modelos de Ollama<\/a> includes Llama 3.1, Mistral 7B, Qwen 2.5, Gemma 2, Phi-3, and dozens more. Each model is available in multiple quantization levels (Q4_K_M, Q5_K_M, FP16) to trade VRAM for quality.<\/p>\n<p>Full installation steps: <a href=\"https:\/\/convly.ai\/es\/how-to-install-ollama-2026\/\">how to install Ollama<\/a>.<\/p>\n<h2><span class=\"ez-toc-section\" id=\"Open-Source_Model_Families_Available_in_Ollama\"><\/span>Open-Source Model Families Available in Ollama<span class=\"ez-toc-section-end\"><\/span><\/h2>\n<table>\n<thead>\n<tr>\n<th>Model Family<\/th>\n<th>Desarrollador<\/th>\n<th>Parameter Sizes<\/th>\n<th>Licencia<\/th>\n<th>Notable For<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>Llama 3.1<\/td>\n<td>Meta<\/td>\n<td>8B, 70B, 405B<\/td>\n<td>Llama 3.1 (commercial-friendly)<\/td>\n<td>Best general reasoning at each size class<\/td>\n<\/tr>\n<tr>\n<td>Mistral<\/td>\n<td>Mistral AI<\/td>\n<td>7B, 22B (Mixtral 8x7B)<\/td>\n<td>Apache 2.0<\/td>\n<td>Fast, efficient, strong coding<\/td>\n<\/tr>\n<tr>\n<td>Qwen 2.5<\/td>\n<td>Alibaba<\/td>\n<td>0.5B, 1.5B, 3B, 7B, 14B, 32B, 72B<\/td>\n<td>Apache 2.0<\/td>\n<td>Multilingual, long context (128k)<\/td>\n<\/tr>\n<tr>\n<td>Gemma 2<\/td>\n<td>Google<\/td>\n<td>2B, 9B, 27B<\/td>\n<td>Gemma (commercial-friendly)<\/td>\n<td>Small, fast, runs on CPU<\/td>\n<\/tr>\n<tr>\n<td>Phi-3.5<\/td>\n<td>Microsoft<\/td>\n<td>3.8B (mini), 14B (medium)<\/td>\n<td>MIT<\/td>\n<td>Compact, strong on reasoning benchmarks<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>El <a href=\"https:\/\/convly.ai\/es\/models\/\">Base de datos de modelos de IA<\/a> lists specs, VRAM requirements, and pricing for 37 models including all of the above.<\/p>\n<h2><span class=\"ez-toc-section\" id=\"Pulling_and_Running_a_Model\"><\/span>Pulling and Running a Model<span class=\"ez-toc-section-end\"><\/span><\/h2>\n<p>Once Ollama is installed, pull a model with <code>ollama pull &lt;model&gt;:&lt;tag&gt;<\/code>. The tag specifies parameter count and quantization. Examples:<\/p>\n<pre><code>ollama pull llama3.1:8b\nollama pull mistral:7b-instruct-q4_K_M\nollama pull qwen2.5:7b\nollama pull gemma2:9b<\/code><\/pre>\n<p>Run the model:<\/p>\n<pre><code>ollama run llama3.1:8b<\/code><\/pre>\n<p>This opens an interactive chat session. Type your prompt, press Enter, and the model generates a response locally. To exit, type <code>\/bye<\/code>.<\/p>\n<p>To use the model programmatically via the API:<\/p>\n<pre><code>curl http:\/\/localhost:11434\/api\/generate -d '{\n  \"model\": \"llama3.1:8b\",\n  \"prompt\": \"Explain recursion in one sentence.\"\n}'<\/code><\/pre>\n<p>Ollama exposes an OpenAI-compatible <code>\/v1\/chat\/completions<\/code> endpoint, so you can point existing OpenAI SDK code at <code>http:\/\/localhost:11434<\/code> and swap in a local model.<\/p>\n<h2><span class=\"ez-toc-section\" id=\"Parameter_Sizes_and_VRAM_Requirements\"><\/span>Parameter Sizes and VRAM Requirements<span class=\"ez-toc-section-end\"><\/span><\/h2>\n<p>Model size (7B, 13B, 70B) determines quality and VRAM footprint. Larger models reason better but need more GPU memory. Quantization (Q4, Q5, FP16) compresses weights to reduce VRAM at a small quality cost.<\/p>\n<table>\n<thead>\n<tr>\n<th>Parameter Count<\/th>\n<th>Cuantizaci\u00f3n<\/th>\n<th>VRAM (approx)<\/th>\n<th>Example Models<\/th>\n<th>GPU Recommendation<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>7B<\/td>\n<td>Q4_K_M<\/td>\n<td>4.5 GB<\/td>\n<td>Llama 3.1 de 8B, Mistral de 7B<\/td>\n<td>RTX 3060 (12 GB), RTX 4060 Ti (16 GB)<\/td>\n<\/tr>\n<tr>\n<td>7B<\/td>\n<td>Q5_K_M<\/td>\n<td>5.5 GB<\/td>\n<td>Same as above<\/td>\n<td>Same<\/td>\n<\/tr>\n<tr>\n<td>7B<\/td>\n<td>FP16<\/td>\n<td>14 GB<\/td>\n<td>Same as above<\/td>\n<td>RTX 4060 Ti (16 GB), RTX 4070<\/td>\n<\/tr>\n<tr>\n<td>13B<\/td>\n<td>Q4_K_M<\/td>\n<td>8 GB<\/td>\n<td>Qwen 2.5 14B, Phi-3.5 14B<\/td>\n<td>RTX 3060 (12 GB), RTX 4060 Ti (16 GB)<\/td>\n<\/tr>\n<tr>\n<td>13B<\/td>\n<td>FP16<\/td>\n<td>26 GB<\/td>\n<td>Same<\/td>\n<td>RTX 4090 (24 GB) or A5000 (24 GB)<\/td>\n<\/tr>\n<tr>\n<td>70B<\/td>\n<td>Q4_K_M<\/td>\n<td>40 GB<\/td>\n<td>Llama 3.1 70B, Qwen 2.5 72B<\/td>\n<td>A100 (40\/80 GB), dual RTX 3090 (48 GB total)<\/td>\n<\/tr>\n<tr>\n<td>70B<\/td>\n<td>FP16<\/td>\n<td>140 GB<\/td>\n<td>Same<\/td>\n<td>Multi-GPU setup or A100 80GB<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>Usa la <a href=\"https:\/\/convly.ai\/es\/llm-vram-calculator\/\">Calculadora de VRAM<\/a> to estimate memory needs for any model and quantization level. The <a href=\"https:\/\/convly.ai\/es\/vram-requirements-every-major-llm-2026\/\">Gu\u00eda de requisitos de VRAM<\/a> lists precise values for every major model.<\/p>\n<p>For GPU buying decisions, see <a href=\"https:\/\/convly.ai\/es\/best-gpus-for-local-llms-2026\/\">mejores GPUs para ejecutar LLMs localmente<\/a>.<\/p>\n<h2><span class=\"ez-toc-section\" id=\"Quantization_Quality_vs_VRAM_Trade-Off\"><\/span>Quantization: Quality vs VRAM Trade-Off<span class=\"ez-toc-section-end\"><\/span><\/h2>\n<p>Quantization reduces model weight precision from 16-bit floats (FP16) to 4-bit or 5-bit integers (Q4, Q5). This cuts VRAM by 60-75% with minimal quality loss for most tasks.<\/p>\n<ul>\n<li><strong>Q4_K_M<\/strong>: 4-bit quantization. Smallest VRAM, slight quality drop on complex reasoning. Good for chatbots, summarization, code completion.<\/li>\n<li><strong>Q5_K_M<\/strong>: 5-bit quantization. ~20% more VRAM than Q4, closer to FP16 quality. Best balance for most users.<\/li>\n<li><strong>Q8_0<\/strong>: 8-bit quantization. Nearly FP16 quality, 50% VRAM reduction. Use if you have the VRAM headroom.<\/li>\n<li><strong>FP16<\/strong>: Full precision. Best quality, double the VRAM of Q4. Only needed for research or if quality regressions are unacceptable.<\/li>\n<\/ul>\n<p>Example: Llama 3.1 8B at Q4_K_M uses 4.5 GB and scores 67.2 on MMLU. At FP16 it uses 14 GB and scores 68.1 on MMLU. For most tasks the 0.9-point difference is imperceptible.<\/p>\n<p>To pull a specific quantization:<\/p>\n<pre><code>ollama pull llama3.1:8b-instruct-q4_K_M\nollama pull llama3.1:8b-instruct-fp16<\/code><\/pre>\n<p>If you omit the quantization tag, Ollama defaults to Q4_K_M.<\/p>\n<h2><span class=\"ez-toc-section\" id=\"Performance_Comparison_7B_Class\"><\/span>Performance Comparison: 7B Class<span class=\"ez-toc-section-end\"><\/span><\/h2>\n<p>The 7B-8B size class is the most popular for local use. It fits consumer GPUs and delivers strong results on coding, reasoning, and chat.<\/p>\n<table>\n<thead>\n<tr>\n<th>Modelo<\/th>\n<th>MMLU (0-shot)<\/th>\n<th>HumanEval (pass@1)<\/th>\n<th>Context Length<\/th>\n<th>VRAM (Q4)<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>Llama 3.1 8B<\/td>\n<td>67.2<\/td>\n<td>62.2<\/td>\n<td>128k<\/td>\n<td>4.5 GB<\/td>\n<\/tr>\n<tr>\n<td>Mistral 7B v0.3<\/td>\n<td>62.5<\/td>\n<td>40.2<\/td>\n<td>32k<\/td>\n<td>4.1 GB<\/td>\n<\/tr>\n<tr>\n<td>Qwen 2.5 7B<\/td>\n<td>70.3<\/td>\n<td>61.6<\/td>\n<td>128k<\/td>\n<td>4.3 GB<\/td>\n<\/tr>\n<tr>\n<td>Gemma 2 9B<\/td>\n<td>71.3<\/td>\n<td>61.0<\/td>\n<td>8k<\/td>\n<td>5.2 GB<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>Qwen 2.5 7B and Gemma 2 9B lead on MMLU (general knowledge). Llama 3.1 8B leads on HumanEval (coding). Mistral 7B is the fastest and most permissive license (Apache 2.0). The <a href=\"https:\/\/convly.ai\/es\/llm-leaderboard\/\">Clasificaci\u00f3n de modelos de lenguaje grande (LLM)<\/a> ranks all models by intelligence, price, and context length.<\/p>\n<h2><span class=\"ez-toc-section\" id=\"Which_Model_to_Pick\"><\/span>Which Model to Pick<span class=\"ez-toc-section-end\"><\/span><\/h2>\n<ul>\n<li><strong>Best overall quality (7B class)<\/strong>: <code>ollama pull qwen2.5:7b<\/code> o <code>ollama pull llama3.1:8b<\/code><\/li>\n<li><strong>Best for coding<\/strong>: <code>ollama pull llama3.1:8b<\/code> o <code>ollama pull qwen2.5-coder:7b<\/code><\/li>\n<li><strong>Fastest inference<\/strong>: <code>ollama pull mistral:7b<\/code> o <code>ollama pull gemma2:2b<\/code> (for CPU\/low VRAM)<\/li>\n<li><strong>Multilingual<\/strong>: <code>ollama pull qwen2.5:7b<\/code> (supports 29 languages)<\/li>\n<li><strong>Long documents (128k+ context)<\/strong>: <code>ollama pull llama3.1:8b<\/code> o <code>ollama pull qwen2.5:7b<\/code><\/li>\n<li><strong>Strongest reasoning (if you have 40+ GB VRAM)<\/strong>: <code>ollama pull llama3.1:70b<\/code> o <code>ollama pull qwen2.5:72b<\/code><\/li>\n<\/ul>\n<p>El <a href=\"https:\/\/convly.ai\/es\/best-local-llms-to-run-on-ollama-2026\/\">mejores modelos de lenguaje locales para Ollama<\/a> guide benchmarks all models and recommends specific tags per use case.<\/p>\n<h2><span class=\"ez-toc-section\" id=\"When_to_Self-Host_vs_Use_an_API\"><\/span>When to Self-Host vs Use an API<span class=\"ez-toc-section-end\"><\/span><\/h2>\n<p>Running Ollama locally makes sense if:<\/p>\n<ul>\n<li>You already own a GPU with 12+ GB VRAM (RTX 3060, 4060 Ti, or better)<\/li>\n<li>You process sensitive data that cannot leave your network<\/li>\n<li>You generate more than 5 million tokens per month (API costs exceed self-hosting TCO)<\/li>\n<li>You need guaranteed uptime and no rate limits<\/li>\n<\/ul>\n<p>Use a hosted API (OpenAI, Anthropic, Groq) if:<\/p>\n<ul>\n<li>You generate fewer than 1 million tokens\/month (self-hosting GPU amortization takes years)<\/li>\n<li>You need the absolute best quality (Claude 3.5 Sonnet, GPT-4o outperform all open models)<\/li>\n<li>You don&#8217;t want to manage inference infrastructure<\/li>\n<\/ul>\n<p>El <a href=\"https:\/\/convly.ai\/es\/self-hosting-vs-api-calculator\/\">calculadora de autohospedaje frente a API<\/a> estimates break-even based on your token volume, GPU cost, and electricity rate. The <a href=\"https:\/\/convly.ai\/es\/ai-api-cost-calculator\/\">Calculadora de costos de API<\/a> projects monthly spend per model.<\/p>\n<h2><span class=\"ez-toc-section\" id=\"Platform-Specific_Notes\"><\/span>Platform-Specific Notes<span class=\"ez-toc-section-end\"><\/span><\/h2>\n<h3>macOS<\/h3>\n<p>Ollama uses Metal for GPU acceleration on M1\/M2\/M3 Macs. Unified memory means VRAM = system RAM. An M2 Max with 64 GB can run Llama 3.1 70B at Q4 (40 GB) with room for the OS. Install via Homebrew:<\/p>\n<pre><code>brew install ollama<\/code><\/pre>\n<p>Start the service:<\/p>\n<pre><code>ollama serve<\/code><\/pre>\n<h3>Linux<\/h3>\n<p>Ollama requires NVIDIA GPUs with CUDA 11.8+ or AMD GPUs with ROCm 5.7+. Install with:<\/p>\n<pre><code>curl -fsSL https:\/\/ollama.com\/install.sh | sh<\/code><\/pre>\n<p>This installs the binary to <code>\/usr\/local\/bin\/ollama<\/code> and creates a systemd service. Start it:<\/p>\n<pre><code>sudo systemctl start ollama<\/code><\/pre>\n<p>Models download to <code>~\/.ollama\/models<\/code>. To change the location, set <code>OLLAMA_MODELS=\/path\/to\/models<\/code> en <code>, y luego ejecuta<\/code>.<\/p>\n<h3>Windows<\/h3>\n<p>Ollama for Windows requires NVIDIA GPUs with CUDA 11.8+ and driver 520+. Download the installer from <code>ollama.com<\/code> and run it. The service starts automatically. Models download to <code>C:Users&lt;YourName&gt;.ollamamodels<\/code>.<\/p>\n<p>To run from PowerShell:<\/p>\n<pre><code>ollama run llama3.1:8b<\/code><\/pre>\n<h2><span class=\"ez-toc-section\" id=\"Frequently_Asked_Questions\"><\/span>Preguntas frecuentes<span class=\"ez-toc-section-end\"><\/span><\/h2>\n<h3>Does OpenAI offer an open-source model I can run in Ollama?<\/h3>\n<p>No. OpenAI released GPT-2 (2019) as open-weight, but all recent models (GPT-3.5, GPT-4, GPT-4o, o1) are proprietary and API-only. If you want OpenAI-quality reasoning locally, try Llama 3.1 70B or Qwen 2.5 72B\u2014both outperform GPT-3.5 on most benchmarks and run in Ollama with 40 GB VRAM at Q4 quantization.<\/p>\n<h3>Can I run Ollama on a CPU without a GPU?<\/h3>\n<p>Yes, but inference is 10-50\u00d7 slower. Small models (Gemma 2 2B, Qwen 2.5 0.5B, Phi-3.5 mini 3.8B) are usable on modern CPUs (16+ threads). Larger models (7B+) will generate 1-3 tokens per second on CPU, which is too slow for interactive use. If you don&#8217;t have a GPU, consider using a hosted API instead\u2014see the <a href=\"https:\/\/convly.ai\/es\/self-hosting-vs-api-calculator\/\">calculadora de autohospedaje frente a API<\/a>.<\/p>\n<h3>How much does it cost to run Ollama compared to OpenAI API?<\/h3>\n<p>OpenAI charges $0.15 per million input tokens and $0.60 per million output tokens for GPT-4o mini. A 12 GB GPU (RTX 4060 Ti, $400) running Llama 3.1 8B costs $0.05\/hour in electricity (at $0.12\/kWh, 400W system draw). Break-even is around 3-5 million tokens\/month. The <a href=\"https:\/\/convly.ai\/es\/ai-api-cost-calculator\/\">Calculadora de costos de API<\/a> y <a href=\"https:\/\/convly.ai\/es\/self-hosting-vs-api-calculator\/\">self-hosting calculator<\/a> show exact TCO for your usage.<\/p>\n<h3>What is the difference between Q4_K_M, Q5_K_M, and FP16?<\/h3>\n<p>Q4_K_M uses 4-bit quantization (smallest VRAM, slight quality loss). Q5_K_M uses 5-bit quantization (20% more VRAM, closer to full quality). FP16 is full 16-bit precision (best quality, double the VRAM of Q4). For most tasks, Q5_K_M is the best balance. Use FP16 only if you have spare VRAM and need the last 1-2% of quality for research or production.<\/p>\n<h3>Can I fine-tune models in Ollama?<\/h3>\n<p>No. Ollama is an inference engine, not a training framework. To fine-tune an open model, use Hugging Face Transformers with PEFT\/LoRA, Axolotl, or LLaMA Factory. Export the fine-tuned weights to GGUF format and import them into Ollama with <code>ollama create<\/code>. El bloque <a href=\"https:\/\/convly.ai\/es\/what-is-ollama-complete-guide-2026\/\">Gu\u00eda completa de Ollama<\/a> covers this workflow.<\/p>\n<h3>Which GPU should I buy to run 7B models locally?<\/h3>\n<p>For Q4 quantization (4.5 GB VRAM): RTX 3060 12 GB ($250 used) or RTX 4060 Ti 16 GB ($450 new). For FP16 (14 GB VRAM): RTX 4060 Ti 16 GB or RTX 4070 ($550-600). For 70B models at Q4 (40 GB): dual RTX 3090 24 GB ($1800 used) or A100 40 GB ($6000+ used). The <a href=\"https:\/\/convly.ai\/es\/best-gpus-for-local-llms-2026\/\">best GPUs guide<\/a> benchmarks price\/performance for every size class.<\/p>","protected":false},"excerpt":{"rendered":"<p>TL;DR OpenAI has not released a model family called &#8220;GPT-OSS&#8221;. The search term likely refers to running open-source alternatives (Llama [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":2200,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"default","site-content-style":"default","site-sidebar-style":"default","ast-global-header-display":"","ast-banner-title-visibility":"","ast-main-header-display":"","ast-hfb-above-header-display":"","ast-hfb-below-header-display":"","ast-hfb-mobile-header-display":"","site-post-title":"","ast-breadcrumbs-content":"","ast-featured-img":"","footer-sml-layout":"","ast-disable-related-posts":"","theme-transparent-header-meta":"","adv-header-id-meta":"","stick-header-meta":"","header-above-stick-meta":"","header-main-stick-meta":"","header-below-stick-meta":"","astra-migrate-meta-layouts":"default","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"ast-content-background-meta":{"desktop":{"background-color":"var(--ast-global-color-4)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"var(--ast-global-color-4)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"var(--ast-global-color-4)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"footnotes":""},"categories":[7],"tags":[],"class_list":["post-2199","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-ai-news"],"_links":{"self":[{"href":"https:\/\/convly.ai\/es\/wp-json\/wp\/v2\/posts\/2199","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/convly.ai\/es\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/convly.ai\/es\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/convly.ai\/es\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/convly.ai\/es\/wp-json\/wp\/v2\/comments?post=2199"}],"version-history":[{"count":1,"href":"https:\/\/convly.ai\/es\/wp-json\/wp\/v2\/posts\/2199\/revisions"}],"predecessor-version":[{"id":2201,"href":"https:\/\/convly.ai\/es\/wp-json\/wp\/v2\/posts\/2199\/revisions\/2201"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/convly.ai\/es\/wp-json\/wp\/v2\/media\/2200"}],"wp:attachment":[{"href":"https:\/\/convly.ai\/es\/wp-json\/wp\/v2\/media?parent=2199"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/convly.ai\/es\/wp-json\/wp\/v2\/categories?post=2199"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/convly.ai\/es\/wp-json\/wp\/v2\/tags?post=2199"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}