Free · Open source · MIT · v0.1.7
Local LLMs and a coding agentthat never leave your machine.
OpenLLM Studio scans your GPU, picks the right model and quantization, then runs chat and a full coding agent offline — private by default. No cloud, no API keys, no telemetry.
New to local LLMs? Read the docs · vs LM Studio · vs Ollama
- Active installs
- 300+ installs
- Telemetry
- 0 bytes phoned home
- License
- MIT licensed
.gif)
Windows · macOS · Linux · CUDA · ROCm · Metal
01 — Detect & Download
Hardware auto-detect.
One-click models.
OpenLLM scans your GPU, RAM, and CPU, then recommends the exact model and quantization that fits your machine. Download GGUF models from Hugging Face in one click — resumable and cached locally.
- NVIDIA CUDA, AMD ROCm/Vulkan, and Apple Metal support
- VRAM-aware model recommendations
- Resumable downloads with local cache
02 — Local Chat
Chat with your models.
Fully offline.
A full chat interface with system prompts, conversation history, and streaming responses. No network required once the model is downloaded — unplug the cable and it keeps working.
- Streaming responses token-by-token
- System prompts and conversation memory
- Works without an internet connection
03 — Coding Agent
An agent that writes code.
On your machine.
Reads, writes, edits, and tests code across an entire project. Runs terminal commands and fixes its own errors — like Cursor or Claude Code, but fully offline and private.
- Multi-file edits with context awareness
- Integrated terminal and code editor
- MCP tool servers for custom integrations
05 — The deal
No subscriptions. No API keys. No telemetry.
No data leaves this machine. That’s not a marketing promise — it’s an architectural fact you can verify. The entire application is MIT-licensed on GitHub. Grep the source for a network call you didn’t ask for. You won’t find one.
$0
cloud costs, forever
100%
offline operation
MIT
every line auditable
06 — Enterprise
Built for teams that can't trust the cloud.
Air-gapped deployments, admin controls, audit logging, and compliance-ready infrastructure — without sending a single token to a third party.
- On-premise / air-gapped installs
- SSO, RBAC, and audit logs
- HIPAA-ready, SOC 2-aligned, GDPR-aware
- Dedicated support & SLA
Admin dashboard
Monitor usage, review audit logs, and enforce policies across every cluster from a single dashboard. No fabricated metrics — real data from your own deployment.
08 — FAQ
Fair questions.
Asked by people who read source code before README badges.
Does it really work fully offline?
Yes. Models run via a bundled inference engine on your GPU or CPU. After a model is downloaded, you can disconnect from the internet entirely — chat, agent, and editor all keep working. The only network activity is the model download you explicitly trigger.
What hardware do I need?
Anything from an 8 GB VRAM GPU upward runs 7B-class models comfortably. The auto-detect picks a model and quantization that fits your VRAM. CPU-only works too, just slower. Apple Silicon runs via Metal out of the box.
How does the agent compare to Cursor or Claude Code?
Same workflow — it reads, writes, edits, and tests code across a project, runs terminal commands, and retries on its own errors. The difference: it uses your local model, so quality tracks the open-source model you run, and nothing you write ever leaves your machine.
Can I really mix NVIDIA and AMD in one cluster?
Yes. Cluster Mode splits model layers across machines over your LAN, and each node uses its own backend — CUDA, ROCm, Vulkan, or Metal. Heterogeneous clusters are a first-class feature, not a hack.
What data do you collect?
None. There is no telemetry, no crash reporting, no analytics, and no account system. If you want to verify that, the source is MIT-licensed on GitHub — search it for outbound requests.
09 — Ship it locally
Your hardware is more capable than you think.
Install in under a minute. Delete it just as fast — nothing lingers, nothing phones home.