Supported Models
Catalog of the 6 AI models supported by AIRGAP Studio
Overview
AIRGAP Studio ships with Qwen3-8B-Q4_K_M.gguf (baseline) bundled in the installer. Additional models can be installed via a separate Modelpack installer or by placing .gguf files directly into the build/vscodium/models/ directory — they will be automatically detected.
The single source of truth for the model catalog is phase3/models-metadata.json, and baseline priority is managed in phase3/model-version.json.
Model Catalog
| Display Name | Filename | Class | GPU Context | CPU Context | Compatibility Tier |
|---|---|---|---|---|---|
| Qwen3 8B (Q4_K_M) | Qwen3-8B-Q4_K_M.gguf | 8B | 40,960 | 16,384 | recommended (default) |
| Qwen3 4B (Q4_K_M) | Qwen3-4B-Q4_K_M.gguf | 4B | 40,960 | 8,192 | compatible |
| Granite 4.0 Micro (Q4_K_M) | granite-4.0-micro-Q4_K_M.gguf | micro | 32,768 | 8,192 | experimental |
| Qwen3 1.7B (Q8_0) | Qwen3-1.7B-Q8_0.gguf | micro | 32,768 | 8,192 | experimental |
| Mistral 7B Instruct v0.3 (Q4_K_M) | Mistral-7B-Instruct-v0.3.Q4_K_M.gguf | 7B | 32,768 | 8,192 | compatible |
| Meta Llama 3.1 8B Instruct (Q4_K_M) | Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf | 8B | 32,768 | 16,384 | compatible |
Compatibility Tier Definitions
- recommended: The officially recommended baseline model for AIRGAP Studio. Optimally verified across all assistants / Designer / Monitor.
- compatible: Verified to work correctly. Response quality on some features may differ from the baseline.
- experimental: Functional but with limited verification of response quality and stability. Intended as a fallback for low-spec environments.
Model Selection Guide
| Environment | Recommended Model | Notes |
|---|---|---|
| VRAM 4GB | Qwen3 1.7B (Q8_0) or Granite 4.0 Micro | Lightweight, fast responses |
| VRAM 6GB | Qwen3 4B (Q4_K_M) | Balanced choice |
| VRAM 8GB or more | Qwen3 8B (Q4_K_M) — default / recommended | Mistral 7B / Llama 3.1 8B also viable |
| CPU-only | Qwen3 1.7B or Granite 4.0 Micro | 16GB+ RAM recommended |
See the System Requirements document for detailed GPU/CPU requirements.
Automatic Model Detection
llama-server automatically recognizes .gguf files in the build/vscodium/models/ directory.
build/vscodium/models/
├── Qwen3-8B-Q4_K_M.gguf # Bundled by default
├── Qwen3-4B-Q4_K_M.gguf # If additionally installed
└── Mistral-7B-Instruct-v0.3.Q4_K_M.gguf
After placing a new model and restarting AIRGAP Studio, it will be added to the model selection UI automatically.
How to Change Models
Model switching is the sole responsibility of the Monitor extension. Other assistants only display the currently active model (read-only).
- Open the Command Palette (
Ctrl + Shift + P) - Run the
AIRGAP: Select LLM Modelcommand - Choose your model from the QuickPick list
- llama-server restarts automatically and loads the new model
For details, see the Monitor documentation.
Model Downloads
Beyond the bundled default (Qwen3 8B), additional models are distributed via a separate Modelpack installer.
- Download: Official downloads page
- For air-gapped environment deployment, simply distribute the Modelpack
.exealongside the main installer.
Related Docs
- System Requirements — GPU/CPU/VRAM requirements
- Compatibility Proxy — Model-specific chat format normalization gateway
- Monitor Extension — Model selection UI