Multi-model · v2.2.0
Run the model you want,
where you want it.
GIDE runs models two ways: fully offline on your own machine, or through a hosted provider with your own API key. Local is private and free to run; hosted gets you frontier quality when you need it. Same editor, same agent. You just pick.
Local models
On-device · GGUF · nothing leaves your machine
Point GIDE at a GGUF file and it handles the rest: memory budgeting, context size, and tool-calling config, all tuned for that specific model. Four families are supported today, and they are genuinely different tools, not the same model in four sizes:
- OrnithDefault
- What ships on by default. Trained for tool-calling, so it drives the agent reliably: reading files, running commands, editing. If you're not sure what to run, run this.
- Qwen
- A strong, well-rounded coder. Handles the same tool-calling workload as Ornith and is the better pick if you already know the Qwen family and like how it writes.
- DeepSeek
- The lightweight one. DeepSeek R1 7B is the only model here that fits comfortably on 8GB, but it has no tool-call training, so the agent is weaker with it. Reach for it when memory is the constraint, not the default.
- Gemma
- Google's Gemma, at 12B the largest of the set. Best output quality when you have the memory for it, and it wants the memory. It won't load on 8GB at all.
Hosted models
Your API key · stored in your system keychain
No local GPU headroom, or you just want the strongest model on a hard problem? Paste an API key and go. GIDE ships no keys of its own. Yours are stored in your OS keychain and only ever sent to the provider they belong to, never to us.
You can also connect NVIDIA OpenShell over MCP. That's not a model; it hands the agent extra tools to call, on top of GIDE's built-in ones.
How much memory you actually need
Mac uses unified memory · Windows uses GPU VRAM
A local model has to fit in memory twice over: the weights, plus the context (KV cache) that grows with every token. More memory buys you a bigger context window. Here's what GIDE will actually give each model, after reserving the ~3.2K-token system prompt and a safety buffer:
| System | Ornith 9B | Qwen 9B | DeepSeek R1 7B | Gemma 3 12B |
|---|---|---|---|---|
| Mac 18GB | 32K | 32K | 64K | 48K |
| Mac 24GB | 64K | 64K | 64K | 64K |
| Win 8GB | 4K(894 usable) | 4K | 24K | won't fit |
| Win 12GB | 32K | 32K | 64K | 32K |
| Win 16GB | 48K | 64K | 64K | 64K |
What we'd actually tell you
On Windows, 8GB VRAM is a bad time. Ornith and Qwen collapse to a 4K window, and after the 3.2K system prompt that leaves under 900 tokens for your file, the diff, and the model's reply. Gemma won't load. The only thing that runs well is DeepSeek 7B, the one model with no tool-call training. You shouldn't get pushed onto a weaker agent just because of a memory floor.
12GB is the real minimum. That's the first tier where every model runs with a workable 32K-plus window, so the model choice is yours, not your GPU's.
Macs have it easy. Unified memory is shared with the model, so 18GB already runs the default at a comfortable 32K, and 24GB takes everything to 64K.
Exact numbers shift with the model file, quantization, and your GPU-offload setting. GIDE picks a safe default per model and you can nudge it in settings.