Install Ollama Local LLM Windows - stone-alex/EliteIntel GitHub Wiki

🪟 Local LLM - Windows Setup (Ollama)

Running a local LLM keeps everything private, offline, and free (beyond electricity and hardware). Think of it as the difference between running a game on your own rig vs. streaming it from the cloud - lower latency, no subscriptions, no one snooping on your loadout.

It requires Ollama and a capable GPU.

Minimum Hardware

To run Elite Dangerous and the LLM on the same machine, you need at minimum an NVIDIA RTX 3060 with 12 GB VRAM. That's the floor - it'll run, but don't expect headroom to spare.

Tip: You can point Elite Intel at an Ollama instance running on a separate PC on your network. If you have a home lab or a spare box with a good GPU, that's a great option - the game PC doesn't carry the load at all.

Recommended Model

Model VRAM Required Notes
tulu3:8b Q4_K_M ~5 GB ✅ Recommended. Reliable for commands and queries.
qwen3 8B ~8 GB Experimental. Expect occasional missed commands and hallucinations.

Note: If you want the fastest local inference, consider LM Studio with matrixportalx/tulu-3.1-8b-supernova. In testing, it's noticeably faster than Ollama on the same hardware with the same model class.

Step 1 - Install Ollama

  • Go to 👉 https://ollama.com/download 👈
  • Download and run OllamaSetup.exe. No admin rights required.
  • Ollama installs and runs quietly in the system tray. It auto-starts on login.

Step 2 - Pull a Model

Open Command Prompt or PowerShell and run:

ollama pull tulu3:8b

Or experimental alternatives:

ollama pull qwen3:8b

Step 3 - (Optional) Tune the Configuration

Out of the box Ollama works fine, but if you want to share VRAM more carefully with Elite Dangerous, this is where you do it.

On Windows, Ollama reads config from user environment variables.

  1. Right-click the Ollama tray icon → Quit.
  2. Open Settings → search "environment variables".
  3. Click "Edit environment variables for your account".
  4. Add each variable below using New:
Variable Value Notes
OLLAMA_MAX_VRAM 14000000000 14 GB cap - adjust based on your GPU and game needs
OLLAMA_NUM_PARALLEL 3 Covers Elite Intel's async call patterns without over-allocating
OLLAMA_MAX_LOADED_MODELS 1 One model at a time
OLLAMA_FLASH_ATTENTION 1 Faster inference
OLLAMA_KEEP_ALIVE -1 Keep model loaded permanently
  1. Click OK. Relaunch Ollama from the Start Menu.

What these settings do

OLLAMA_MAX_VRAM - hard cap on VRAM Ollama can use, in bytes. Leaves the rest for Elite Dangerous. Adjust based on your GPU and how much the game needs.

OLLAMA_NUM_PARALLEL - how many requests Ollama handles simultaneously. Elite Intel makes async calls so setting this too low causes failures. 3 covers the typical command + query overlap without over-allocating.

OLLAMA_MAX_LOADED_MODELS - keeps only one model in VRAM at a time. No reason to keep stale models loaded.

OLLAMA_FLASH_ATTENTION - enables Flash Attention, which reduces memory bandwidth usage during inference. Generally faster, especially for repeated requests.

OLLAMA_KEEP_ALIVE=-1 - keeps the model loaded in VRAM indefinitely. Without this, Ollama may unload the model after a period of inactivity and you pay a reload penalty on the next request.

Step 4 - Wire It Up in Elite Intel

Head to the Settings tab in Elite Intel:

  • Leave the LLM Key field blank (local Ollama doesn't need one).
  • LLM Address defaults to http://localhost:11434/api/chat. If Ollama is on another machine, replace localhost with that machine's IP.
  • LLM Model - set to tulu3:8b.
  • Command LLM - set to tulu3:8b.
  • Query LLM - set to tulu3:8b.
  • Hit Stop → Start on the AI tab to apply changes.

Community 👉Matrix👈