Build your own AI.
Run it privately.
ModelLM is the control layer above the local AI ecosystem. It helps you find a model, work out whether your hardware can actually run it, add your own knowledge, and decide whether you need retrieval, fine-tuning, both, or neither — then trains, evaluates and ships it.
- — chat app
- — model repository
- — inference engine
Find → Understand → Customize → Train → Evaluate → Optimize → Deploy → Run
- 01
Base model
Pick weights that fit your machine
- 02
Knowledge
Index what you know
- 03
Dataset
Clean what you will train on
- 04
Train
LoRA / QLoRA on your GPU
- 05
Evaluate
Prove it beat the base model
- 06
Optimize
Quantize to fit and to fly
- 07
Deploy
Ollama, llama.cpp, local files
- 08
Run
Use it, privately
One model. Different modes.
A mode is a goal, not a setting. Pick what you are trying to achieve and ModelLM decides which techniques that implies — retrieval, a fine-tune, quantization, tools — and which it can talk you out of.
Knowledge
Give a model access to your own documents.
Training
Teach a model a behaviour, format or domain style.
Performance
Make a model fit, and make it fast.
Private
Nothing leaves the machine.
Agent
Give a local model tools and let it act.
Coding
Build a local coding assistant.
Personality
Shape voice, format and refusal behaviour.
Private + Knowledge + Training + Agent
= a private company agent
Private company agent
An internal assistant that knows your documents, answers in your format, can act on internal systems, and never sends a byte outside the building.
Local coding AI
A completion and review model quantized to sit permanently in VRAM alongside your editor, with your source never leaving the machine.
Research assistant
Retrieval over a paper library with citations back to the source PDF. No training, no GPU required, minutes to build.
Support responder
Trained on resolved tickets for structure and tone, retrieving live product documentation for the facts.
Can I run this?
ModelLM sizes every model against your actual machine: weights at each quantization, the KV cache for the context you want, and the runtime overhead nobody accounts for. It tells you what fits, what fits with offloading, and what will not run at all.
These are arithmetic from published model geometry — and ModelLM says so, every time. It will not invent a tokens-per-second figure it has not measured.
23 GB usable VRAM
ModelLM will talk you out of fine-tuning.
Most people who think they need to train a model need retrieval instead. ModelLM analyses what you actually uploaded — how many documents, how many examples, how good they are, how often they change — and then tells you which of the four approaches fits, with the evidence behind it.
You can always disagree. The button that says fine-tune anyway is right there.
RAG
Recommended because:
- only 47 documents
- information changes frequently
- knowledge should remain editable
- no instruction dataset to train on
Illustration of the advisor panel
This is where you build your own AI.
ModelLM ships with a working demo workspace, a model catalogue and a hardware compatibility engine. Nothing to configure, nothing uploaded anywhere.