Run Qiplim Studio and any LLM (cloud or local) on your own hardware. BYOK, model-agnostic, your infrastructure, your rules.
Looking for a managed cloud option instead? Discover Pando Studio (qiplim.com/studio) — hosted in France, Mistral by default, GDPR-compliant. See the managed cloud option
Nothing leaves your network unless you decide. Documents, generations, and student data stay where you put them.
Plug in Mistral, OpenAI, Anthropic, Google, or a local Ollama instance. No vendor lock-in, no per-token markup.
Swap providers anytime. Run a local 7B for low-stakes tasks and a hosted model for harder ones. You stay in control.
No network latency, no rate limits, no outages from external providers. Works offline.
A 7-8B parameter model is sufficient for generating quizzes, summaries, flashcards, and podcasts with good quality.
| Configuration | RAM | Supported models | Est. performance | Approx. price |
|---|---|---|---|---|
| Mac Mini M4 (base) | 16 GB | 7-8B: Mistral Small 3, Qwen 3 8B, Llama 3.3 8B | ~30-35 tok/s | ~700 EUR |
| Mac Mini M4 Pro | 24 GB | 8-14B: Qwen 3 14B, Gemma 3 12B | ~25-30 tok/s | ~1,200 EUR |
| Mac Mini M4 Pro | 48 GB | 32B: Qwen 3 32B, Mistral Medium | ~10-15 tok/s | ~1,800 EUR |
| Refurbished PC (i7/Ryzen, GPU 8GB+) | 32 GB | 7-13B with GPU acceleration | ~15-25 tok/s | ~300-500 EUR |
Power consumption: Mac Mini draws 30-40W under AI inference load (~10-15 EUR/year electricity).
Ollama manages local LLM models and exposes an OpenAI-compatible API.
macOS (Homebrew)
brew install ollama
ollama serveLinux
curl -fsSL https://ollama.com/install.sh | sh
ollama serveChoose a model that fits your hardware. 7-8B models work well on 16GB RAM.
ollama pull mistral-small
# or: ollama pull qwen3:8b
# or: ollama pull llama3.3:8bGet the source code (early access) and start the services.
git clone REPO_URL_ACCES_ANTICIPE/studio.git # contactez-nous pour l'acces anticipe
cd studio
pnpm install
pnpm docker:up # PostgreSQL + Redis
pnpm db:push
pnpm devSet your database, Redis, and AI provider settings in .env.
cp .env.example .env
# Edit .env:
DATABASE_URL=postgresql://qiplim:qiplim@localhost:5433/qiplim_studio
REDIS_URL=redis://localhost:6379
BETTER_AUTH_SECRET=your-secret-here
BYOK_ENCRYPTION_KEY=your-32-char-key-hereMake the studio accessible to other devices on your LAN.
# In .env, set your LAN IP:
NEXT_PUBLIC_APP_URL=http://192.168.1.42:3001
# Start the production server:
pnpm build && pnpm startNative support for custom base URLs (Ollama, vLLM, LocalAI) is coming in the next release. Qiplim Studio will let you point any provider to a local endpoint directly from the settings.
In the meantime, you can use LiteLLM as a proxy to expose Ollama as an OpenAI-compatible endpoint:
pip install litellm
litellm --model ollama/mistral-small --port 4000LiteLLM translates requests to/from the Ollama API. Qiplim Studio sees a standard OpenAI endpoint.
Upload PDF, DOCX, PPTX, TXT, CSV files directly via the Studio interface or the API.
Connect a Google Drive folder to automatically import and index documents.
Use POST /api/documents/upload to integrate with your existing content pipeline.
Once running, access Qiplim Studio from any device on your network:
Find your server's local IP (e.g. 192.168.1.42) or use mDNS hostname (e.g. mac-mini.local)
Set NEXT_PUBLIC_APP_URL=http://192.168.1.42:3001 in your .env
Access from any browser on the network at http://192.168.1.42:3001
Optional: set up a reverse proxy (Caddy or nginx) for HTTPS with a self-signed certificate
Computer labs, university departments, language labs. Generate quizzes and interactive content from course materials without sending student data to the cloud.
Internal training, onboarding, compliance. Process confidential documents (HR policies, internal procedures) on your own infrastructure.
Qiplim Studio is preparing its open-core release. Contact us for early access and deploy on your own hardware.