Llama.cpp setup: GGUF models, GPU tuning and OpenCode

This Spanish-language tutorial explains how to run models locally with Llama.cpp using precompiled Windows builds. The speaker uses Windows 10 and an NVIDIA RTX 3060 with 12 GB of VRAM, downloads the CUDA build and its libraries, and organizes executables and GGUF model files in separate folders. CPU execution and other GPU backends also receive a brief explanation.

Model selection depends on available memory, including the space needed for context. The walkthrough uses quantized Qwen 3.5 9B files from Hugging Face and adds a vision projector. A Windows .bat file launches llama-server with a browser chat interface. The speaker adjusts sampling settings, reasoning and MTP speculative decoding, recommending tests on the same prompt rather than assuming a draft-token setting will always improve speed.

When the configuration exceeds GPU memory, the tutorial explores Q8 KV cache quantization and moving vision processing to CPU. It also configures Gemma 4 E2B with a separate draft model, then splits Qwen 3.6 35B A3B experts between GPU and system RAM. The speaker reports usable generation speeds but warns that processing long inputs is slower with CPU offloading.

The final sections connect the OpenCode coding assistant to an OpenAI-compatible local endpoint and briefly explain presets.ini. These presets let llama-server unload one model and load another without separate launch files.