# How Mirakl uses Anyscale

**3x**: “saved the company 3x in inference costs compared to using OpenAI API for all listings”

As published on [anyscale.com](https://www.anyscale.com/resources/case-study/mirakl-ai-infrastructure). Captured by usedby on 2026-10-07.

- Company: [Mirakl](https://www.usedby.ai/companies/mirakl.md)
- Tool: [Anyscale](https://www.usedby.ai/tools/anyscale.md)
- Industry: [Marketplace](https://www.usedby.ai/companies/industry/marketplace.md)
- Teams: data engineers, data scientists

## What the story says

Mirakl uses Anyscale's managed Ray platform to run fine-tuned, smaller open-source LLMs with LoRA adapters for batch inference. This automates catalog onboarding for marketplace sellers, with GPU clusters that autoscale to match traffic.

Summary written by usedby from the source page, in English. The figures are those of Anyscale and Mirakl, not ours.

> Investing in fine-tuned open-source models and leveraging Anyscale to manage Ray orchestration saved the company 3x in inference costs compared to using OpenAI API for all listings.

- **80k+ token/sec**: “Consistent peak performance, handling 80k+ token/sec generations without latency.”
- **10 million**: “automating onboarding for over 10 million products each month”
- **20 GPU nodes**: “they can scale up to 20 GPU nodes during peak hours, but scale down to near-zero GPUs overnight”
- **90%**: “By intelligently routing up to 90% of catalog traffic to LLaMA 3.1 8B”

> Using GPT-4 would have cost us more than half a million dollars per month… with Anyscale and fine-tuned LLaMA models, we brought that down by 3× while scaling to millions of products.
>
> Arthur Delaitre, Manager of Data Science

## What usedby checked

We compared the story with its live page on 2026-10-07.

- Checked: the figure 3x is printed word for word on the page, near the name of Mirakl.
- Checked: the passage quoted above is copied word for word from the page, near the name of Mirakl.
- Checked: Mirakl uses Anyscale. Confirmed line. Latest check across sources: 2026-10-07.
- Not checked: the result itself. We quote it; we did not measure it.

---
Source: https://www.usedby.ai/case-studies/mirakl-anyscale · How we check: https://www.usedby.ai/methodology