Pricing

Lamini On-Demand currently uses MI250s, but we have MI300s available for our Lamini Reserved plans. Please contact us to learn more about Lamini Reserved and our MI300 cluster.

Increasing the number of GPUs will speed up your job by approximately 1.5x per GPU. Lamini will automatically reschedule your long running jobs, even if they’re only scheduled on 1 GPU.

For Lamini On-Demand, the price for both input and output tokens is $0.50 per million tokens.

Not for Lamini On-Demand. If you want to run a large volume of jobs or data, contact us about Lamini Reserved or Self-managed for better pricing.

For Lamini Reserved and Self-Managed, we license based on the number and type of GPU(s). Please contact us for a quote.

Yes, we do. Please contact us.

For an initial evaluation data set, you will need about 20-40 input-output pairs to start. As you iterate, you will add more data until you achieve the level of accuracy required for your use case.

It takes approximately 50 steps for every 100 data points you want to train, but this will vary significantly based on size and complexity of your data points. We calculate tuning job costs by: $1 per step * number of GPUs. Example: Memory tuning 100 data points with 50 steps → $50 on one GPU or $50 * 2 = $100 on 2 GPUs

In the context of tuning models, a "step" refers to a single update of the model's weights / one iteration. You can set the number of steps you want per job when you submit it.

Yes! Our free $300 in credits is enough to run the Meta Llama Notebook and tuning jobs from scratch.

Yes, if you created an account earlier, you should have received $300 in free credit. If you didn’t receive your credit, please contact us.

You can request more GPUs for your job. Each additional GPU will improve performance by about 1.5x. Requesting more GPUs will increase the cost of the job.

We built our inference engine to be highly performant. We run on AMD MI250 and MI300 GPUs and Nvidia H100 GPUs so our Single Stream memory wall is 200 tokens/sec, 331 tokens/sec, and 209 tokens/sec respectively. Learn more about evaluating performance of inference frameworks here.

A datapoint is a single instance of data used in training. For example, in a text classification task, each sentence or document would be a datapoint. The number of datapoints affects the overall training time and cost.

Steps are provided by the user when submitting a job. By default, we assume 50 steps per 100 datapoints, but this can be adjusted based on your specific needs. More complex tasks or larger models might require more steps per datapoint.

Using multiple GPUs can significantly speed up the training process. Each additional GPU provides approximately a 1.5x speed increase, allowing you to train your model faster. This can be particularly beneficial for large datasets or complex models.

This calculator provides a rough estimate based on typical usage. Actual costs depend on the number of steps and GPUs provided.