NEURAX is theLovable of models.
Build your AI. Run it your way.
Memory, speed and cost in milliseconds, without a GPU, an account or a cluster.
Available for Linux · macOS and Windows coming soon
The studio, in your browser
Change the design.Watch what it costs.
Pick a model, edit a block, choose a machine and run the analysis. Every figure below was produced by the NEURAX compiler; nothing is drawn by hand.
Loading the studio…
Every number, computed before anything runs.
- Compiles your design into operations: never runs it to find out
- Local design and training; optional Agent requests go to the provider you choose
- Works with Hugging Face and the AI provider you already use
Works with the frameworks, providers and hardware you already use
- Hugging Face
- OpenAI
- Anthropic
- Gemini
- DeepSeek
- OpenRouter
- GLM
- NVIDIA
- AMD
- Intel
- macOS
- Linux
- Windows
Workflow
One model, one environment, one loop
Every stage feeds the next, and the last one feeds the first. The second half happens on real hardware; that half is what closes the loop.
01 · computed
Design
The model, block by block
Drag blocks onto the canvas, load a published architecture, or import a Hugging Face config.
02 · computed
Analyze
Parameters, FLOPs, tensor shapes
The compiler lowers every block to primitive operations and counts them exactly.
03 · computed
Simulate
Memory, speed, cost
Peak memory against what your machine really has free, step time, training hours and spend.
04 · computed
Plan
Batch, precision, schedule
Sweep batch sizes, precisions and device counts to find the setting that fits.
05 · on real hardware
Train
A real run on real hardware
NEURAX writes the PyTorch project, checks it matches the analysis, and runs it on your machine.
06 · on real hardware
Observe
Loss, throughput, memory, live
Every step reported as it happens, with checkpoints so a run can stop and resume.
07 · on real hardware
Compare
Predicted against observed
What NEURAX predicted, beside what the run actually did.
08 · on real hardware
Optimize
The next design, better informed
Change the design with the measured run in hand, and go round again.
NEURAX Assistant
Justdescribethe model
Bring your own key. If you use the optional Agent, your chosen provider receives its prompt and project design/data profile; NEURAX says it does not send the raw dataset or local path. Provider terms apply.
Example questions for the NEURAX assistant: Size a model that fits a 24 GB card for ticket triage; Why is the attention block memory-bound here?; Cut peak VRAM by a third without dropping layers
Your key: OpenAI · Anthropic · Gemini · Mistral · DeepSeek · OpenRouter · GLM · Local endpoint
Held to PyTorch itself
The compiler's ground truth is 73 published architectures measured with PyTorch: parameters, FLOPs by operator, and the bytes autograd keeps. An automated test compiles 56 of them on every change.
0
published architectures measured with PyTorch
0
compiled by an automated test on every change
≤0%
parameter drift allowed before the test fails
- gpt-oss-120bMixture of experts116,829,156,672+0.00%
- Mixtral-8x7BMixture of experts46,702,792,704+0.00%
- Qwen3-30B-A3BMixture of experts30,532,122,624+0.00%
The test fails if the compiler drifts more than 1% on parameters or 5% on FLOPs. Models it does not yet reproduce are listed with the reason, and that list is only allowed to shrink.
Who it is for
Built for people who pay for compute
Where it runs
- Workstationavailable
- Local serveravailable
- Cloud GPUplanned
- AI clusterplanned
01
AI research
Evaluate architectures before committing compute.
02
ML engineering
Understand training requirements before execution.
03
MLOps
Run workloads against real infrastructure, not a spec sheet.
04
Universities
Help teachers teach AI visually and practically: students see a model take shape and what it costs.
05
FinOps
Put a defensible number on every workload before it runs.
06
Engineering leadership
Make infrastructure decisions from measured data.
No more pasting architectures into chat windows.
Connect Claude Desktop, Cursor or any MCP client, and it can analyze an architecture, estimate training cost and check a budget with NEURAX's own engine.
Works with any MCP client.
Questions, answered
Do I need to know machine learning?
No. You need to know your problem and your data. Choosing an architecture, sizing it, writing the PyTorch and fitting it to your hardware is the work NEURAX does.
Does it run my model to measure it?
No. It compiles the architecture to the operations it's made of and prices them, which is why it answers in milliseconds, and why it can cost a model far larger than your machine.
Does my data leave my machine?
The Studio reads supported datasets locally for profiling and training. If you use the optional Agent, your chosen provider receives its prompt and project design/data profile; NEURAX says it does not send raw dataset contents or the local path. Importing a public Hugging Face or ModelScope model also makes an online request for its configuration.
Do I need a GPU?
Not to design, cost or compare models: that runs on a CPU. You need one to train quickly, and NEURAX tells you what a run costs on the machine you have, including a laptop.
Can it really train, or only plan?
It trains. The run is written into your project folder as real files, runs on your machine, reports every step live, and checkpoints so it can be paused and resumed.
What do I own at the end?
Ordinary PyTorch files and trained weights, in your own folder. Nothing is locked to NEURAX.
Can we use our own internal LLM?
Yes. The assistant talks to whatever endpoint you configure, including a model hosted inside your own network.
What does it cost?
Community is free. Developer is $4.99 a month and Pro is $20 a month; Enterprise is priced with you. The assistant runs on your own provider key, so its model usage is whatever your provider charges.
Know before you spend a GPU-hour
Free to download. No account. Runs on the machine in front of you.