PricingWorkspace·Open workspace
Free tier includes starter gems and 5 GB storage

Vibe-build, train and tune your own AI models

WhisperAI Terminal is a multimodal model trainer and creator. Describe what you want a model to do and the Forge console builds it end to end — gathering the data, choosing the base model, training, distilling, refining and testing. Or take the controls yourself: start from scratch, from a Hugging Face import, or from a model or dataset already in your storage, on the GPU tier you choose, with resources you monitor and every checkpoint yours to download.

From an idea to a trained model, at any starting point

The Forge console is a vibe-style build console for models. Nothing but a sentence is required to begin, and every stage stays visible, priced and under your control.

  1. 1

    Say what it's for

    "A support assistant that knows my help site" or "shrink this 70B teacher into something I can run on one GPU" is enough to start.

  2. 2

    Get a plan and a price

    The Forge picks the smallest capable base model, lists every step — crawl, import, initialise, fine-tune, distill, clean-up — and shows the gem cost and runtime before anything runs.

  3. 3

    Start from any level

    From nothing at all (design the architecture yourself), from an open-weight or catalog model, from a Hugging Face import, or from a model, adapter or dataset already in your storage.

  4. 4

    It builds and refines

    Data is gathered, runs are queued on the GPU tier that fits, and later passes can refine the checkpoint the earlier pass produced. Temporary sources are deleted when it's done.

  5. 5

    Test, then download

    The finished model is checked against acceptance tests written for your goal — ask it anything yourself — then download the weights and keep them.

Vibe-build a model

Describe the model you want in plain English. The Forge console writes the build plan, prices it up front, runs every step and tests the result.

Pull from the hubs

Search Hugging Face models and datasets, the Ollama library and GitHub, then import straight into your workspace.

Crawl your own data

Point the crawler at any site, pick which media types to harvest, preview pages in the tuning browser, and it becomes a JSONL dataset.

Train, distill or tune

Fine-tune a catalog model, distill a big teacher into a small student, or keep training a model already in your storage — full control of method, precision, LoRA rank, epochs, batch size, learning rate, steps and seed.

Metered compute

Choose a GPU tier and see the gem price and estimated runtime before you start. Cancel mid-run and unused time is refunded.

Bring your own files

Upload weights, adapters and datasets from your device into private per-account storage.

Yours to keep

Every model, checkpoint and dataset is downloadable and deletable at any time.

Compute tiers

Billed per minute in gems — the same wallet as WhisperAI Bots. Rates track what the GPU host actually costs on Fly.io (100 gems = $1), so you pay close to metal plus a small operating margin.

AcceleratorMemoryGems / minHost costBest for
CPU pool8 GB RAM1$0.23/hr · performance-4x · 8 GBTiny models, embeddings, tokenizer work
L40S48 GB5$1.53/hr · l40s · performance-4x · 16 GBLoRA on 7B, small vision runs
A1024 GB5$1.86/hr · a10 · performance-8x · 32 GBComfortable 7–13B fine-tuning
A100 40 GB40 GB8$2.86/hr · a100-40gb · performance-8x · 32 GBFull fine-tunes, 30B LoRA, diffusion
A100 80 GB80 GB HBM2e12$4.28/hr · a100-80gb · performance-8x · 64 GBFrontier-scale training and distillation

Job types

Fine-tune

Adapt a base model to your dataset (LoRA or full)

Model → model distill

Teach a small student model from a larger teacher

Embed / index

Turn a dataset into vectors for retrieval

Crawl & build dataset

Harvest text and media from the web

Plans

Free tier: 5 GB storage, free daily gems, one run at a time.

Lab

$9.99/mo

or $99.90/yr — 2 months free

  • 50 GB model & dataset storage
  • 600 gems every month
  • 10% off all compute
  • Up to 2 concurrent runs
  • Accelerators up to A10
Choose Lab

Forge

$24.99/mo

or $249.90/yr — 2 months free

  • 250 GB model & dataset storage
  • 1,800 gems every month
  • 25% off all compute
  • Up to 4 concurrent runs
  • Accelerators up to A100
Choose Forge

Foundry

$59.99/mo

or $599.90/yr — 2 months free

  • 1 TB model & dataset storage
  • 5,000 gems every month
  • 40% off all compute
  • Up to 8 concurrent runs
  • Accelerators up to H100
Choose Foundry

Extra storage add-ons

Stack as many as you need — each bills monthly.

+10 GB

$3.45/mo

+30 GB

$9.20/mo

+50 GB

$13.80/mo

+100 GB

$23.00/mo

Get the WhisperAI Terminal app

Install WhisperAI Terminal to your phone, tablet or desktop — its own icon, full screen, no browser bar.

On desktop: click the install icon in the address bar, or use the browser menu → Install WhisperAI Terminal.

Installs directly from this page — no app store account needed.

Sloe the WhisperAI sloth icon

Sloe trains a model

a very short comic
Panel 1Sloe the WhisperAI sloth lounging with a glowing laptop
"Small dataset. Clean rows."
Panel 2Sloe the WhisperAI sloth hugging a soundwave speech bubble
Loss goes down. Sloe goes horizontal.
Panel 3Sloe the WhisperAI sloth waving hello
"It hums back. Ship it."

Loss curve going down

Sloe the WhisperAI sloth lounging with a glowing laptop

serotonin going up

"It's just a small dataset"

Sloe the WhisperAI sloth hugging a soundwave speech bubble

it beat the big one

GPU: on fire

Sloe the WhisperAI sloth waving hello

Sloe: horizontal

Sloe the WhisperAI sloth waving hello

Sloe can help

  • Start with a small, clean dataset before a huge messy one.
  • Watch the loss curve like a kettle — a steady drop is good.
  • Bundle specialists instead of one model doing everything.
Sloe the WhisperAI sloth lounging with a glowing laptop

Sloe's note on training

Sloe teaches the young sloths by repeating a single sound until they hum it back. Training a model is the same patient loop — small lessons, many times.

  • A clean, small dataset beats a huge messy one.
  • Watch the loss curve like a kettle — steady drop is good.
  • Bundle specialists instead of asking one model to do everything.