01
Guided first run
Checks your machine, picks a model size that fits it, helps you get text, and walks you to a working model. Every step can be skipped.
Version 2.0 — rebuilt from the ground up
Train your own language model, from scratch, on your own machine.
Haiku Studio runs the whole pipeline locally: build a tokenizer from your text, pretrain a model, fine-tune it to hold a conversation, align it on the replies you liked, and chat with the result. Nothing is sent anywhere. There is no API key.
Active project
my-first-model
Model49.5M params
Corpus1.3B tokens
Budget50× per param
Pretraining — step 18,400 of 63,700
loss 2.41 · 41,200 tokens/sec
ConsoleTraining
0001 LOG [cache] 1,341,882,304 tokens (3.87 bytes/token)
0002 LOG [eval] step 18000 val 2.44 ppl 11.5 ** new best
0003 WARN [plan] 2.1 passes over corpus at this budget
Capabilities
It tells you the truth about what you are doing.
Most of the work in training a small model is knowing whether a run is worth starting. Haiku Studio answers that before you commit hours to it, not after.
02
Honest planning
Before a run starts you see the model size, the tokens it needs, the passes over your corpus that implies, and whether it can work at all.
03
Environment check
Catches the classic trap where pip install torch silently gives you the CPU-only build, before you lose a week to it.
04
Tokenizer builder
Learn a byte-level BPE vocabulary from your own corpus, with proper control tokens so replies begin and end cleanly.
05
The full pipeline
Pretraining, supervised fine-tuning and DPO alignment, with resumable checkpoints. Stop any run and it picks up exactly where it left off.
06
Chat Lab
Streaming replies from your model, with a stop button that actually stops. Rate replies with thumbs and they become alignment data.
07
Dataset catalog
Download prepared corpora straight into your project. Checksum-verified, and resumable if your connection drops mid-transfer.
08
Runs where you tell it
Pick CPU, a specific GPU, or Apple Silicon. Asking for a device that is not there raises an error instead of quietly using the CPU.
09
A real command line
Everything the app does is available as python -m haiku.cli. The desktop app is a front end, not a wall.
Model sizes
Start small enough to actually finish.
Five presets, from something that trains on a laptop CPU in minutes to a model that will take a good GPU several days. The app recommends one based on your hardware.
Preset
Parameters
Text it wants
Realistic on
Nano
2.3M
~180 MB
Any laptop CPU
Toy
6.5M
~500 MB
CPU (slow) or any GPU
Small default
49.5M
~4 GB
6 GB+ GPU
Medium
127.6M
~10 GB
12 GB+ GPU
Large
393M
~31 GB
24 GB+ GPU, days of training
“Text it wants” is the compute-optimal ratio of roughly 20 tokens per parameter. You can train well past it, and for small models you probably should — the budget dial goes to 200×, which is where small models start to feel capable. Higher ratios need proportionally more text and time, and the app computes both before you commit.
Workflow
Everything a model needs
Everything a model needs
lives in one project folder.
Your project
projects/my-first-model
Corpus, tokenizer, config, checkpoints, cache and logs. Switching projects switches all of them together.
Before you run
plan & readiness
Text present, tokenizer built, budget checked. Training will not start until the prerequisites are actually met.
While it runs
resumable checkpoints
Stop whenever you like. Resume restores the optimizer, the random state and the exact sample it left off at.
Interface
Designed for visibility, not decoration.
Every tab maps to a real stage of the engine, and every number on screen comes from the run itself rather than an estimate.
Data & Tokenizer
Corpus files3 files · 5.2 GB
Total tokens1.34B measured
Tokenizer16,000 vocabulary
Budget dial
3× demo to 200×
Good to go
Console
0001LOG[cache] train 1,331,204,096 · val 200,000 tokens
0002LOG[eval] step 18000 val 2.44 ppl 11.5 ** new best
0003WARN[plan] only 2.1 passes over corpus at this budget
0004ERR[device] cuda requested but PyTorch sees no GPU
Pretraininglearns language
Instructlearns to answer
Alignmentlearns your taste
Chat Lab
Replies stream token by token from a model held in memory, so the first word arrives in milliseconds rather than seconds.
The alignment loop
- Chat with the model you just trained.
- Rate replies with thumbs up and down.
- Those ratings become preference pairs.
- Run DPO and the model learns your taste.
Download
Get Haiku Studio 2.0.
Download the release, or clone the repository and run the Python engine directly. Both give you the same thing — the desktop app is a front end over the command line, not a separate product.
No Windows installer yet. 2.0 ships as source; a packaged installer is coming. You will need
Python 3.10+ and Node.js LTS. Install PyTorch first from
pytorch.org —
a plain pip install torch gives you the CPU-only build on Windows, even with an NVIDIA card.
The built-in environment check will tell you which one you actually have.