Langtrain Logo
Langtrain
PricingDocsSecurity
Sign inStart training
  1. Before you start01
  2. How it works02
  3. Ask it again03
  4. Why not ChatGPT04
  5. Your data05
  6. Models06
  7. Use cases07
  8. Products08
  9. Pricing09
  10. FAQ10
NEWlangtrain-ai 1.0: Langtune and Langvision in one Python SDK

Your own ChatGPT. Private, offline, yours.

Langtrain turns open-source models like Llama, Qwen and Mistral into an AI assistant trained on your company's data. You get the model weights to serve behind an OpenAI-compatible API, on your own hardware, even with no internet once it's trained.

Start training freeRead the quick start
Sensitive data or a security review? Talk to us
  • Weights are yours to keep
  • Runs offline once trained
  • MIT-licensed libraries
  • Train on your GPU or ours

Member of NVIDIA Inception, NVIDIA's program for startups.

trainingEXAMPLE RUN
llama-3.1-8b · LoRA r=16
TRAIN LOSS
0.590
STEP
0520 / 1200
EPOCH
1.30 / 3
  1. step 0400loss 0.693 lr 1.5e-4
  2. step 0450loss 0.631 lr 1.4e-4
  3. step 0500loss 0.599 lr 1.3e-4
Langtrain Logo
Langtrain

Fine-tune open models on your data. Keep the weights.

  • Github
  • HuggingFace
NVIDIA Inception Program memberNVIDIA Inception Program member

Member of the NVIDIA Inception programme for startups.

Product

  • Langtrain Studio
  • Langtrain SDK
  • Langtune
  • Langvision
  • Model Hub
  • Fine-tuning
  • Evaluations
  • Downloads
  • Pricing
  • Enterprise

Use Cases

  • Customer Support AI
  • Internal Code Assistants
  • Healthcare
  • Financial Services
  • Legal Document QA
  • E-Commerce & Retail
  • Education & EdTech
  • Manufacturing
  • Research & Data Teams
  • All Use Cases

Resources

  • Documentation
  • Quick Start
  • API Reference
  • Python SDK
  • Node SDK
  • ComparisonsNew
  • GlossaryNew
  • Changelog

Company

  • About Us
  • Team
  • Careers
  • Press
  • Contact
  • Community
  • Support

Trust

  • Security
  • Data processing agreement
  • System status
  • Cookie policy
Langtrain

© 2026 QUADTREE AI TECHNOLOGIES PRIVATE LIMITED. Built in Hyderabad.

  • Terms of Service
  • Privacy Policy

Before you start

What you bring. What you get back.

The full quick start
  1. 01 / You bring
    • Example conversations as a JSONL file, one per line
    • A Hugging Face login, for gated models like Llama
  2. 02 / You train on
    • Your own NVIDIA GPU: an 8B model fits on 24 GB with QLoRA
    • Or Langtrain's GPUs, from your account
  3. 03 / You get back
    • A standard Hugging Face model folder, yours to keep
    • That llama.cpp, vLLM and Ollama run, behind an OpenAI-compatible API

How it works

Five steps to your own AI. On your GPU or ours.

Upload what your team already has: help articles, tickets, chats, documents. Langtrain turns it into training data, trains an open model on it, checks it, and hands it back to run wherever you want.

Read the quick start
01 Data02 Train03 Evaluate04 Own05 Serve
  1. 01DataUpload chat-format JSONL. One conversation per line.
  2. 02TrainLoRA, QLoRA, DoRA or preference tuning, on your GPU or ours.
  3. 03EvaluateScore the fine-tune on held-out data before it ships.
  4. 04OwnKeep the weights and run them on your GPU, offline with llama.cpp, or on your servers.
  5. 05ServeServe it with vLLM behind an OpenAI-compatible API.

Ask it again

A generic model guesses. Yours knows.

It has never seen your tickets, contracts or policies. Fine-tune an open model on them, then ask the base model and yours the same questions and read the answers side by side.

Read the evaluation guide
Illustrative example · same prompt, two models

Question: Can I get a refund on my annual plan after 30 days?

Generic model(wrong)

Most companies offer refunds within 30 days of purchase. After that, refunds usually aren't available.

Your fine-tuned model(correct)

Yes. Annual plans are pro-rated after day 30, so you get back the months you haven't used.

Why Langtrain

Why not just use ChatGPT or Claude? Because it isn't yours.

Hosted assistants are a good start. They stop working for you the moment your data is sensitive, your domain is specialised, or your costs have to be predictable.

Langtrain compared with hosted AI APIs and building your own stack
CapabilityLangtrainChatGPT / Claude APIsBuild it yourself
Your data stays on your hardwareYesNoYes
Trained on your company's dataYesLimitedYes
You own the model weightsYesNoYes
Works with no internet once trainedYesNoYes
Changes only when you retrain itYesNoYes
No training stack to buildYesYesNo
OpenAI-compatible APIYesTheir ownYou build it
How you payFree to start, then a planPer token, for as long as you use itGPUs and engineers

Hosted APIs change their offerings often; check their current terms.

Where your data goes

Your data stays yours. So does the model.

Train on your own GPU and nothing is uploaded. Train on ours and the weights still end up in a repo you own. Either way you get standard weights that run offline.

  • 01 / PATH

    Train on your own GPU

    Langtune fine-tunes on your machine. Save with merge_before_save=True and you have full model weights in a folder.

    Your data leaves your machine
    Never
    Library
    Langtune (MIT)
    $ pip install "langtrain-ai[train]" langtune
  • 02 / PATH

    Train on Langtrain's GPUs

    Your dataset is uploaded to Langtrain for the run. When it finishes, push the merged weights to a Hugging Face repo you own.

    Your data leaves your machine
    For the run
    Weights go to
    Your repo, private by default
    $ curl -X POST https://api.langtrain.xyz/api/v1/training/jobs/$JOB_ID/export -H "X-API-Key: $LANGTRAIN_API_KEY" -H "Content-Type: application/json" -d '{"repo_id": "you/my-assistant"}'
  • 03 / PATH

    Run it offline

    Either way, convert the weights with llama.cpp and run the model on a laptop or server that never goes online.

    Needs internet
    No
    Tool
    llama.cpp (open source)
    $ python convert_hf_to_gguf.py ./my-assistant --outfile my-assistant.gguf

No certifications yet: we have not completed a SOC 2 or ISO 27001 audit. For regulated data, train in your own environment.

  • Security
  • Data processing agreement
  • System status

Models

Start from open models. Leave with your own.

Browse the model hub

Good first models

  • Qwen2.5 1.5B Instruct1.5B

    Small enough to train quickly. The fastest way to see your data change a model.

    Licence · Apache 2.0

  • Phi-3.5 mini Instruct3.8B

    Still small enough for one GPU, with noticeably better answers.

    Licence · MIT

  • Qwen2.5 7B Instruct7B

    The size most teams ship. QLoRA fits it on a single 24 GB GPU.

    Licence · Apache 2.0

Use cases

Built for teams whose data can't leave. One platform, every department.

All use cases
  • Customer Support AIAnswers from your help centre and past tickets, in your tone.
  • Internal Code AssistantsA coding assistant trained on your private repositories.
  • HealthcareClinical text processed inside your own environment.
  • Financial Services AIFilings, reports and client data analysed in-house.
  • Legal Document QAPrivileged documents searched without any outside API.
  • E-Commerce & RetailProduct and order questions answered in your brand's voice.
  • Education & EdTechTutors trained on your own course material.
  • Manufacturing & IndustrialManuals and maintenance logs, answered on the shop floor.
  • Research & Data TeamsFine-tunes you can reproduce, compare and defend.

Products

A desktop app, an SDK, a CLI and an API. Use whichever suits your team.

  • StudioDesktop app for managing datasets and following your fine-tuning runs.Windows · Linux
  • SDKPython and Node client libraries to start runs from your scripts and notebooks.Python · Node
  • CLIUpload data, start runs and follow their logs from the terminal or a CI job.Terminal · CI
  • APIREST API for datasets, fine-tuning runs and your account.REST

Pricing

Start free. Pay as your model earns its keep.

Compare plans

  • Starter

    Your first private model.

    $0forever

    • 5M training tokens
    • Langtrain Studio (desktop)
    • You own the weights
    Start training free

  • Pro

    For teams training and serving models.

    $20per month · ₹1,499 in India

    • 10 agents
    • Langtrain Studio Pro
    • 200M training tokens
    • Browser agent sessions
    See Pro

  • Enterprise

    For data that must stay in your environment.

    Customannual agreement

    • Train in your own environment
    • Dedicated GPUs
    • Contract and signed DPA
    Talk to us

FAQ

Before you start. Straight answers.

More answers in the full FAQ.

How is this different from ChatGPT or Claude?
Those are general assistants you rent through someone else's servers. Langtrain gives you your own: an open-source model fine-tuned on your data, with weights you keep and can run on your own hardware, including fully offline.
Can it really run offline?
Yes. Train with Langtune on your own GPU, or push a cloud run's weights to your own Hugging Face repo. Then convert the weights with llama.cpp and run the model on a laptop or server with no internet connection.
Which models can I start from?
Open models such as Llama, Qwen, Mistral and Phi. The model hub lists every base model you can fine-tune.
Do I need ML expertise?
Not to start. The SDK can look at your dataset and suggest a model and training setup, and the defaults work for most runs. You will need to run Python on a machine with an NVIDIA GPU, or train on ours.
What data format should I use?
JSONL, CSV or a Hugging Face dataset. The recommended format is JSONL with a messages array per line, the same chat format shown above.
How much data do I need?
Quality matters more than quantity. 50–100 examples adapt style and tone, 500–1,000 are recommended for a specific task, and 10,000+ build domain expertise.
How long does training take?
It depends on the model size, how much data you have and the GPU. Every run streams its progress, so you can see how far along it is.
Is my data private?
If you train with Langtune on your own GPU, your data never leaves your machine. For cloud runs, the dataset is uploaded to Langtrain so it can be trained on; the privacy policy explains how it is stored and handled.
Can I export and self-host my model?
Yes. Train locally with Langtune and the weights are on your disk, or push a cloud run's merged weights to your own Hugging Face repo. They are standard weights, so llama.cpp, vLLM and other open-source tools can run them.

Your data.Your model.

$pip install "langtrain-ai[train]" langtune

Building it yourself? Train your first private model free and keep the weights. Sensitive data or a security review? Talk to us before you upload anything.

Start training freeTalk to us