Skip to content
NEURAX
Download
Try itRun the studio right here, on real engine output

NEURAX is theLovable of models.

Build your AI. Run it your way. A local AI model Studio for university labs, companies and independent builders: design, simulate, estimate resources, run native Training or supported LoRA Fine-tuning, and export.

Download for free

Desktop app for Linux, macOS and Windows

BERT-base · NEURAX Fine-tuning
Current NEURAX Fine-tuning Studio capture: a pinned, read-only BERT-base architecture prepared for Banking77's 77 classification labels
The pinned google-bert/bert-base-uncased model source in NEURAX, with its revision, Apache-2.0 licence and Safetensors inventory

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
  • Design, data profiling and supported runs stay on your machine; optional Agent requests go to the provider you choose
  • Works with Hugging Face and the AI provider you already use
BERT-base · Fine-tuning architecture
The imported BERT-base architecture in NEURAX Fine-tuning, shown read-only with 77 Banking77 output classes
Architecture pinned

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

From model design to export, in one Studio

Analyze and estimate first, review the Decision for the workflow you chose, then run locally and follow the results.

  1. 01 · computed

    Design

    The model, block by block

    Drag blocks onto the canvas, load a published architecture, or import a Hugging Face config.

  2. 02 · computed

    Analyze

    Parameters, FLOPs, tensor shapes

    The compiler lowers every block to primitive operations and counts them exactly.

  3. 03 · computed

    Simulate

    Memory, speed, cost

    Peak memory against what your machine really has free, step time, training hours and spend.

  4. 04 · computed

    Time Machine

    Separate Training and Fine-tuning projections

    Compare multi-year scenarios from the assumptions you enter; projections do not cover inference or production serving.

  5. 05 · on real hardware

    Decision

    Review the proposal before each run

    Check the distinct criteria and approval for native Training or LoRA Fine-tuning before anything starts.

  6. 06 · on real hardware

    Run locally

    Training or Fine-tuning on your machine

    Run a canvas design or adapt a supported published checkpoint with LoRA after its proposal is approved.

  7. 07 · on real hardware

    Compare

    Predicted against observed

    Review the estimate next to the measurements reported by the completed run.

  8. 08 · on real hardware

    Export

    Metrics, checkpoints and final files

    Follow the run in the relevant Live view, then inspect its export files, verification and provenance.

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

This machine · 5 GB freeDataset profile

Your key: OpenAI · Anthropic · Gemini · Mistral · DeepSeek · OpenRouter · GLM · Local endpoint

Held to PyTorch itself

The v1.2.0 release sweep checked all 80 shipped templates and ran 23 real PyTorch forward/training checks. Exact compiler counts are tested separately from runtime smoke checks; neither proves a model will fit or learn well on every client machine.

73

published architectures measured with PyTorch

56

compiled by an automated test on every change

≤1%

parameter drift allowed before the test fails

ModelParameters, PyTorchNEURAX
  • 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%

Automated checks reject parameter drift above 1% or FLOP drift above 5%. The published report lists the test scope and limits; measured results do not guarantee task quality or compatibility with a client's hardware.

Who it is for

For labs, teams and people building independently

Where it runs

  • Workstationavailable
  • Local serveravailable
  • Cloud GPUplanned
  • AI clusterplanned

  • 01

    University AI labs

    Compare supported architectures and run bounded research pilots on local compute.

  • 02

    Students and educators

    Explore model structure, compare resource needs and learn through supported workflows.

  • 03

    Company ML teams

    Inspect a proposal and its resource estimate before a local run.

  • 04

    Independent builders

    Design and analyze personal projects, then use supported local Training or LoRA Fine-tuning.

  • 05

    Researchers and educators

    Review assumptions, approval criteria and run evidence for a study or lesson.

  • 06

    Compute owners

    Compare local workload estimates before allocating a workstation or starting a run.

NEURAX MCP server

Ask any assistant if it fits

Read the docs

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

  • Is NEURAX only for companies and research labs?

    No. Labs and company ML/R&D teams are core audiences, and individuals can use the same Studio for learning, research and personal projects. Start with a template or explore a model visually; training still depends on a supported model, prepared data and compatible local hardware.

  • Do I need to know machine learning?

    You can explore templates, model structure and compute tradeoffs without being an expert. For a run, you still need a supported model and prepared data; NEURAX helps you inspect the design and local requirements before you decide.

  • 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?

    Yes. Native Training runs an approved canvas design. Fine-tuning separately adapts supported published checkpoints with LoRA. Both run locally and report run-specific metrics and artifacts; each has its own Decision approval.

  • What do I own at the end?

    Native Training produces its project files and weights in your project folder. Fine-tuning exports the LoRA adapter and configuration, a merged Safetensors model, verification and provenance into client storage.

  • 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 to download now. Developer is $4.99 a month and Pro is $20 a month as indicative USD prices; paid subscriptions are not open yet. Enterprise is priced with you. Ask for a demo to discuss a plan. The assistant runs on your own provider key, so its model usage is whatever your provider charges.

Build a model that fits your work and your machine

Explore the complete Studio, from design and analysis to approved local runs and export. Free to download. No account.

Download for freeRequest a demo