How do I begin?
Run three system checks, train your first model in Develop, then meet the first customer’s targets in Operate. The next goal at the top takes you where you need to go.
Free · no installation or sign-up
How to play and frequently asked questions
Start with one small model. Build your own AI lab.
Cold Terminal is a free local LLM idle game about training models and growing an AI lab through subscriptions and API services.
This is a simulation: it does not download real models or call paid APIs. Model performance, funds and revenue are game values.
Getting started
Start in Terminal
Run three system checks, then use prompts to earn starting funds. You can also press Space.
Train in Develop
Choose a base model, method and training examples for the customer’s request. Compare format, evidence, speed and memory alongside quality.
Find customers in Operate
Six customers return with new problems. Deliver a model that meets their targets, collect rewards and unlock automation while earning subscription and API income.
Reinvest in hardware and models
Add GPUs, data pipelines and parts fabricators. Expand your lab to gain permanent research and unlock larger models.
What can you do?
- Models and training
- 25 base models · 3 specialties: code, documents and chat · 5 training methods
- Terminal
- 648 prompts · cron automation · Easter eggs paying 1.5–100 times the base reward
- Service operations
- Customer quality, speed and memory targets · subscriptions and API sales · queues and lab automation
- Long-term progress
- 6 chapters · 24 customer jobs and 24 expansions to reach the ending · 12 optional challenges
- Saves and offline progress
- Browser autosave every 5 seconds · 8 hours of offline progress, upgradeable to 48
Frequently asked questions
Is this real training or money?
This is a local AI studio simulation. Model performance, prices and revenue are game values. No real models are downloaded and no paid APIs are called.
Do small models stay useful?
Customers have different quality, format, evidence, speed and memory targets. Smaller tuned models can suit tight hardware or fast responses; larger models help with expansion and teacher knowledge.
How do models improve?
Choose a method and examples. LoRA trains small adapters; full fine-tuning updates all weights. DPO aligns answer preferences and distillation creates a lighter student. Quality carries over within each specialty, base model and example focus, with up to six runs per method.
Why can’t I deliver after meeting targets?
Your lab also needs enough memory to load the model. Check the missing conditions beneath the model selector. Before model routing unlocks, delivery also launches that model as a service.
When does automation unlock?
One customer job unlocks queues, four unlock automatic launch and eight unlock budgeted investment. Enable automation to prepare suitable training. Your queued jobs go first and wait until their resources and memory are available.
Why is automation waiting?
Operate shows the resource it is waiting for. Quick, balanced and quality policies keep 20%, 30% and 40% of funds in reserve. Buy needed hardware yourself while automatic investment is locked or off.
Why did more hardware not increase income?
Revenue comes from customers and completed requests. Hardware previews show income and training effects. Promote services or enable APIs when capacity is spare; try smaller models or quantization when memory is tight.
What changes with quantization?
4-bit quantization reduces weight memory. In this game it lowers quality by two points. The launch comparison lets you switch each model between 8-bit and 4-bit to compare targets.
Should I always charge more?
Higher prices increase revenue per customer but reduce demand. New prices apply to new subscriptions and renewals. APIs bill completed input and output tokens. Sales focus chooses server shares for subscriptions and APIs.
What are parts for?
Parts support larger-model training, servers, clusters, NAS and optimizations. Fabricators consume funds, included in the top income figure.
What carries over after expansion?
Funds, hardware and subscribers restart. Models, customer history, queues, discoveries, permanent research and automation settings remain. New requests and larger base models unlock. Automatic expansion must be enabled separately.
What if I choose the wrong direction?
Change it free anytime in Lab records. Efficient models, API business and document specialization offer different current bonuses. Permanent research earned through expansion stays.
Does it progress while closed?
Returning applies income, training, deliveries, production, cron, queues and automation for up to 8 hours, extendable to 48 with NAS and expansion. Only unlocked and enabled automation runs; events wait for your choice.
Where is progress saved?
Progress saves to this browser every five seconds. Export or import a save in More. Private browsing and clearing site data can remove progress.
What about cron and Easter eggs?
Use Unlock in Terminal and follow the notice for automatic prompts. Special responses pay 1.5 to 100 times the normal reward. Discovered recipes are kept in Lab records.
How do I reach the ending?
Complete 24 jobs for six returning customers and expand the lab 24 times. Track progress in Lab records. The 12 additional challenges and long-term achievements are optional; play continues after the ending.