Skip to content

Context Forge — local models

Context Forge is the local model engine built into Context Code: it downloads and runs GGUF models straight on your machine, with no external server to install or configure. Ideal for working offline, with full privacy and no paid tokens.

Open it from the sidebar: Context Forge (“Local engines and model catalogue”).

The first thing Forge does is get ready for your machine: it downloads the engine binaries (llama.cpp builds) that match your hardware — CPU, and GPU acceleration where available (CUDA for NVIDIA, Vulkan, and other variants depending on the platform). Only the engines you pick get downloaded, from the Engines tab.

The models tab is the Forge Hub — GGUF Models, with two routes:

  • Curated catalog: a ready-to-use selection; pick the model and hit Download.
  • Search HuggingFace: type a term (“gemma 4”, “qwen coder”) or paste an exact author/model repository to see the available GGUFs.

For each model you choose the quantization — versions of the same model with a different balance between size/speed and quality. Rule of thumb: smaller quantizations fit in less RAM/VRAM and answer faster, at the cost of some accuracy; every model ships with a sensible default quantization.

Downloads can be cancelled and resumed later (they pick up where they left off).

Forge Hub: curated catalog, Hugging Face search and per-model quantizations.

Under Catalogue → Installed, hit Run on a model to load it into memory/GPU (and Stop to unload it and free RAM/VRAM). With the model running, “Context Forge” shows up as just another provider in the chat selector.

The app’s top bar has shortcuts to free memory (unload the local model) or stop the local engine entirely.

Forge exposes the running model through an OpenAI-compatible API, so you can point any tool that speaks that protocol at it:

  • Base URL: the one shown in the API tab (served on 127.0.0.1, reachable only from your own machine).
  • API key: a new one is generated every time the server starts; copy it from that same tab. Every request carries the Authorization: Bearer <API key> header.
  • The model field in your requests must be the id of a running model (the tab lists the “Loaded models”).
Ventana de terminal
curl <Base URL>/v1/chat/completions \
-H "Authorization: Bearer <API key>" \
-H "Content-Type: application/json" \
-d '{"model": "<id del modelo cargado>", "messages": [{"role": "user", "content": "Hola"}]}'
The API tab: OpenAI-compatible Base URL, key and loaded models.
Context Forge downloads the engine your machine needs (CUDA, ROCm, Vulkan or CPU) and keeps it up to date.
The multimedia catalog adds image, video and voice to the text models.
The Design view uses those multimedia models: image, video and voice generated on your own machine.