Language models

Run a chat assistant on your own hardware, read long documents, and process thousands of rows of text without sending anything to an API.

SOC 2 controls 142 verified providers 99.2% uptime Ready in about 90 seconds

What do you need to do?

Private chat assistant Like the ones you use, on a machine only you can reach. Summarize long documents Contracts, reports, transcripts, in one pass. Extract fields from files Invoices and forms into a spreadsheet. Classify at volume Thousands of tickets sorted by your own categories. Translate Long documents, formatting kept intact. Work through hard problems A model that shows its reasoning.

Models for this work

All categories
Popular Llama 3.3 70B A capable general assistant that runs entirely on your own hardware. from $1.35/hr about 34 seconds per thousand words Fastest Mistral Small 3.1 A quick, cheap assistant that also reads images. Good default for volume work. from $0.45/hr about 24 seconds per thousand words Best quality Qwen2.5 72B The strongest open model for reasoning, math, and non-English work. from $1.35/hr about 36 seconds per thousand words Gemma 3 27B A mid-sized assistant that reads images and handles very long documents. from $0.89/hr about 29 seconds per thousand words DeepSeek-R1 Distill 70B Thinks through a problem before answering. Slower, and better on hard questions. from $1.35/hr about 55 seconds per question

Which one should I pick?

ModelBest forQualitySpeedPrice
Llama 3.3 70B Private work with documents you are not allowed to send to an API Excellent Moderate $1.35/hr
Mistral Small 3.1 High volume text jobs where speed and cost matter more than the last few points of quality Very good Very fast $0.45/hr
Qwen2.5 72B Reasoning, math, structured extraction, and work in languages other than English Excellent Moderate $1.35/hr
Gemma 3 27B Long documents and mixed text-and-image work at a middling price Very good Fast $0.89/hr
DeepSeek-R1 Distill 70B The small number of problems where being right matters more than being quick Excellent Slow $1.35/hr

Prices are the hourly rate for the hardware we recommend for each model, billed by the minute. Every model page has a calculator that turns that into a total for your job.

Two people working together at a laptop
NO SURPRISES

You will never get a surprise bill.

  • A hard spend cap. New accounts start at $25. Nothing starts that would go past it.
  • Auto-shutoff when idle. Sessions stop after 15 minutes, with a warning at 12.
  • A live cost meter. Spent so far and budget left, on screen the whole time.
COMING SOON

Renting opens shortly. Want to know when?

We are testing the rental flow with a small group before we open it to everyone. Leave an email and we will tell you the day it opens. One message, no newsletter.

Questions people ask

Why run a language model here instead of using a commercial API?

Privacy and price. Nothing you send to your session leaves the machine we start for you, and the machine and its disk are destroyed when you stop. That is what makes it usable for contracts, medical notes, HR files, and customer records that you are not permitted to send to a third party. On price, a session costs between $0.45 and $2.25 per hour no matter how much you send through it, which works out cheaper than per-token pricing once you are processing at volume.

Are these as good as the commercial assistants?

Close, and the gap is smallest on the work most people actually do. For summarizing, extraction, classification, rewriting, and translation, the large open models are hard to tell apart from the commercial ones. The gap is still real on long multi-step reasoning and on very recent world knowledge. If the task has a right answer and it is hard, use Qwen2.5 72B or DeepSeek-R1. If the task is writing or processing, any of them will do.

How much text can I give it at once?

Between about 250 and 300 pages, depending on the model. Gemma 3 and Qwen2.5 both hold roughly 128,000 tokens, which is a long annual report or a stack of contracts. That entire amount stays in memory, so the model can answer questions about page 3 and page 240 in the same breath. Going past the limit does not error, it quietly forgets the earliest part, which is why we size the recommended card to fit the full context.

How much does it cost to process 10,000 documents?

Roughly $4 to $12, depending on document length and the model. Mistral Small 3.1 on an RTX 4090 at $0.45 per hour handles about 150 short documents per minute, so 10,000 is a little over an hour. A batch job shuts the GPU off the moment the last row is done, so you pay for the processing and nothing else. The calculator on each model page gives you a figure before you commit.

Can I use these models in a product I sell?

Usually yes, and we show the license on every model page. Mistral Small 3.1 is Apache 2.0, which has no conditions at all. Llama 3.3 permits commercial use unless your product has over 700 million monthly users, and requires a "Built with Llama" credit. Qwen and Gemma have their own terms that permit commercial use with some restrictions. If you are building a product, start with an Apache 2.0 model and you will never have to revisit the question.

What happens to my documents after the session ends?

Files you save into your workspace folder stay in your account. Everything else goes: the machine is terminated and its disk is wiped. Batch job inputs and outputs are deleted from our storage 24 hours after completion. We do not read, store, or train on the contents of your sessions.

Own a GPU that sits idle? Put it to work. The average provider earns $180 to $420 per month per card.

List your GPU